Distributed anomaly detection with attention-guided diffusion models and client-side defect generation

1 Distributed Anomaly Detection with Attention-guided Diffusion Models and Client-side Defect Generation Pasquale Coscia , Senior Member, IEEE, Angelo Genovese , Senior Member, IEEE, Vincenzo Piuri , Fellow, IEEE, Konstantinos N. Plataniotis , Fellow, IEEE, and Fabio Scotti , Senior Member, IEEE Abstract—Modern industrial systems are increasingly defined by geographically distributed production lines, stringent privacy constraints, particularly to protect intellectual property and manufacturing process details, and heterogeneous data pipelines. In such environments, centralized Anomaly Detection (AD) is often impractical due to data governance restrictions and limited computational resources at local sites. To address these challenges, we propose a modular and lightweight AD framework based on diffusion models, named D-ADDA (Distributed Anomaly Detection based on Data Augmentation), designed for distributed deployment. Unlike many state-of-the-art methods that depend on large pre-trained models or external datasets, our approach is trained entirely on defective data locally available, enhancing privacy and domain specificity. A novel Data Augmentation Module (DAM) generates diverse defective samples through a multi-stage pipeline, which are used to train an attention-based diffusion model for defect synthesis. This architecture supports dislocated components across multiple clients, enabling training and inference in resource-constrained or privacy-sensitive settings. Experimental results on the MVTec AD dataset confirm the effectiveness of our approach, achieving an average classification accuracy of 60.46% across 14 categories, outperforming state-ofthe-art approaches, with competitive localization performance. Our code will be made publicly available upon publication. Index Terms—Distributed anomaly detection, diffusion process, data augmentation, attention mechanism, defect synthesis. I. INTRODUCTION INDUSTRIAL systems are rapidly evolving into complex, distributed infrastructures comprising geographically dispersed production sites, heterogeneous sensing systems, and tightly regulated data governance frameworks. Achieving reliable performance and high production quality in decentralized settings requires not only intelligent monitoring but also scalable and privacy-preserving solutions to protect details of the industrial process [1]. Localized intelligence at the point of data acquisition is therefore becoming increasingly critical, P. Coscia, A. Genovese, V. Piuri, and F. Scotti are with the Department of Computer Science, Universit` a degli Studi di Milano, 20133, Milan, Italy. (e-mail: [email protected], [email protected], vin- [email protected], [email protected]). K. N. Plataniotis is with the Department of Electrical and Computer Engineering, University of Toronto, M5S 3G4, Toronto, ON, Canada. (e-mail: [email protected]). This work was supported in part by the EC under projects EdgeAI (101097300) and GLACIATION (101070141), and by project SERICS (PE00000014) under the MUR NRRP funded by the EU - NGEU. Project EdgeAI is supported by the Chips Joint Undertaking and its members including top-up funding by Austria, Belgium, France, Greece, Italy, Latvia, Netherlands, and Norway under grant agreement No. 101097300. Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union, the Chips Joint Undertaking, or the Italian MUR. Neither the European Union, nor the granting authority, nor Italian MUR can be held responsible for them. as it avoids transmitting sensitive information such as defect patterns or process information. In such contexts, practical constraints such as latency, limited network availability, and stringent requirements for preserving the confidentiality of sensitive information, particularly defective or anomalous samples protected by industrial secrecy or regulatory mandates, pose additional challenges to the design and deployment of Machine Learning (ML)-based solutions. These limitations undermine the viability of traditional centralized training and model deployment paradigms, which assume unrestricted access to comprehensive datasets. Among industrial tasks, Anomaly Detection (AD) aims to detect deviations from expected patterns to identify system failures, operational inefficiencies, or defects [2]. AD is particularly challenging in manufacturing sites, especially for ML-based approaches, since operational data is abundantly available, while defective samples are scarce, highly sensitive, and often subject to strict confidentiality constraints [3], [4]. In such scenarios, collaborative strategies like distributed learning become essential, enabling different sites or devices to jointly train ML-based AD models without sharing raw data. However, when data types (e.g., normal vs. defective) are unevenly distributed across nodes, this imbalance can lead to biased models, hinder convergence during training, and degrade overall detection performance, particularly for rare but critical anomalies. In fact, while deep learning models have shown promise in AD [5]–[7], they still face limitations related to data quality, training instability, domain shift, and generalization [8]–[10]. Alternative strategies like knowledge distillation, embedding-based methods [11], [12], and multimodal extensions [13] have attempted to address these issues but the reliance on scarce defective samples and pre-trained models often constrains performance. In response to operational demands across modern industrial systems, we propose a modular architecture for image-based AD that supports deployment in environments with limited computational resources and strict privacy requirements, especially to protect confidential intellectual property and proprietary process information. Our method leverages generative diffusion models, trained locally on defective data augmented through a novel pipeline, and supports distributed execution across distinct clients or devices. Instead, a central server processes data that can be safely shared, keeping confidential and proprietary information local, thereby accelerating training and improving quality inspection (see Figure 1). More specifically, we present a three-stage approach built on a lightweight diffusion-based architecture (48.3 M parameters, compared to the 860 M parameters of Stable Diffusion v1.4 [14]), trained © 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. 2 Fig. 1: System architecture considered in this work, showing local nodes (e.g., production lines) connected to a central server for privacy-preserving anomaly detection. The numbered steps (1–7) indicate the operational workflow. exclusively on collected data without relying on large-scale pre-trained models, thereby enhancing adaptability while addressing both privacy and resource constraints. Our diffusion-based generative model, named D-ADDA (Distributed Anomaly Detection based on Data Augmentation), leverages a large set of defective samples created through a data augmentation pipeline, which simulates defects similar to those in the training data, thereby increasing the model’s awareness of the defective domain. The defective samples are processed only at a local level, while the central server only handles non-sensitive data. Our framework prioritizes the protection of defective data, as these samples are most critical for AD and directly expose production failures. However, in industrial practice, even non-defective data may also be sensitive, since it can reveal proprietary or novel production processes. Thus, while the focus remains on defective data, sensitivity considerations can extend to both defective and non-defective samples depending on the confidentiality requirements of the application context. When local-only processing is possible and sufficient computational resources are available, running the full pipeline, including augmentation, generative synthesis, and classifier or detector training, for each domain or object type can still improve accuracy. Therefore, the system not only enables privacyaware deployment, but also provides a performance-enhancing strategy. Beyond its technical contributions to AD, D-ADDA also provides broader systems engineering benefits. By integrating diffusion-based generative modeling with attention-driven feature alignment in a distributed framework, D-ADDA provides a novel contribution to industrial image analysis and also directly supports industrial system-level objectives. In particular, its lightweight design and privacy-preserving operation enhance fault tolerance, reduce downtime, and improve adaptability of monitoring infrastructures under real-world constraints. Moreover, our approach demonstrates adaptability to different categories, making it suitable across diverse industrial contexts. Our main contributions are as follows: •A modular architecture that allows the core components of the AD pipeline, such as data augmentation, synthesis, detection, localization, and classification to operate independently across different locations. This supports incremental deployment across new production sites and facilitates scalability and maintainability in distributed environments. •A decentralized anomaly detection approach that enables efficient, scalable, and privacy-preserving infrastructures by using centralized resources only for training on normal samples or masks, without privacy leaks. •We outperform similar architectures on the MVTec AD dataset, demonstrating the effectiveness of the proposed solution. The remainder of this paper is organized as follows. Section II reviews prior research on anomaly detection and generative approaches. Section III provides a detailed description of each stage of the proposed system architecture. Section IV reports extensive experimental evaluations, including ablation studies on key parameters of the framework. Finally, Section V summarizes the main contributions of the work and outlines potential directions for future research. II. RELATED WORK Anomaly Detection. AD deals with discovering deviations from expected patterns and generally encompasses classification, detection, and localization tasks, with the main challenge being the scarce availability of data describing anomalous situations. To overcome this issue, unsupervised methods address detection and localization by relying solely on normal samples, using either features extracted from deep neural networks or reconstruction methods such as autoencoders, diffusion models, and image inpainting [15], [16].Within unsupervised learning, recent advancements have shown promising results using zero-shot [17], [18] and few-shot approaches [19], [20]. For example, MAEDAY [21] uses a reconstruction-based approach with a pre-trained masked autoencoder to identify anomalous regions as areas of failed reconstruction. CLIP [22] also describes an unsupervised approach that provides strong generalization capabilities through its joint alignment of textual and visual features. Building on this foundation, CLIPbased approaches aim to further enhance this alignment for anomaly detection [23], [24]. To deal with privacy concerns, Chen et al. [25] propose a decentralized approach to mitigate overgeneralization in unsupervised reconstruction; nevertheless, the method does not fully eliminate the necessity of exchanging sensitive information, including defect-related data. Federated learning techniques have also been proposed to mitigate datasharing concerns [25], [26], but they introduce communication overhead and potential privacy risks [27]. In fact, although a number of approaches have advanced privacy-preserving anomaly detection, each presents inherent inefficiencies that limit their practical applicability to the industrial domain. For example, lattice-based frameworks [28] introduce considerable computational overhead arising from fine-grained feature-level policy negotiation, while edge-computing solutions [29] rely on lightweight encryption to protect data but still incur substantial latency and communication overhead, rendering them 3 unsuitable for latency-sensitive IoT environments. Similarly, the bi-level federated learning framework [30] enhances data privacy and scalability; however, it suffers from repeated transmission of encrypted model updates and continued dependence on manual labeling of anomalies in the second training stage, both of which hinder efficiency and scalability in industrial deployments. In this regard, our framework maintains privacy and removes the dependency on data encryption by focusing on the exchange of only non-sensitive data. In fact, sensitive samples are processed locally, achieving an efficient balance between collaboration and confidentiality. Anomaly Generation. Given the scarcity of anomalous samples, generative models can help by creating additional data or capturing the distribution of normal (or abnormal) behavior. While model-free methods, such as cropping and pasting textures [31], can introduce synthetic anomalies, they often lack realism and fail to capture the subtle characteristics of real defects. In contrast, model-based approaches [6] produce more lifelike anomalies that better reflect real-world distributions. From an architectural standpoint, Generative Adversarial Networks (GANs) were among the first to be adopted for this task. For example, OCR-GAN [32] decomposes images into multiple frequency bands and reconstructs them in parallel, modeling the AD process as a form of omni-frequency restoration. However, despite their early success, GANs suffer from well-known issues such as training instability [9]. More recently, Denoising Diffusion Probabilistic Models (DDPMs) have emerged as a promising alternative, demonstrating impressive capabilities in generating high-quality images across diverse domains [10], [33]. Innovations like compressed latent spaces [14] and spatial conditioning mechanisms [34] have further enhanced their expressiveness and ability to model complex data distributions. In this regard, AnomalyDiffusion (AnomDiff) [7] leverages latent DDPMs to generate high-fidelity synthetic anomalies. By including dual encoders to disentangle defect appearance and location, along with an attention-based module to highlight subtle anomalies, AnomDiff improves the alignment between generated masks and image content, although challenges remain in fully replicating real defective images. Similarly, DRÆM [35] generates pseudo-anomalous images to train a U-Net-based [36] reconstruction framework, which learns to map abnormal inputs back to their normal counterparts. A discriminator further refines this process by distinguishing between pseudo-anomalous and reconstructed outputs, aiding in the precise localization of defects. UniAD [37], on the other hand, employs Transformers [38] with masked selfattention to reconstruct features while introducing a feature jittering strategy to prevent shortcut learning. To combine the advantages of generating both text and images, multi-modal approaches offer improved alignment between supervisory signals and enable better discrimination between visually similar defects, even if these methods often entail higher computational costs. For instance, Lee and Choi [13] propose a text-guided AD framework that aligns semantic priors from text with normal image representations, thereby enhancing the selection of discriminative samples. Nonetheless, such methods typically depend on carefully engineered prompts and pre-trained models, which can restrict adaptability and generalizability. In contrast to existing approaches, our method does not rely on pre-trained architectures, thus allowing greater flexibility in the design of custom models. To mitigate data scarcity, we first apply an advanced data augmentation pipeline to enrich the available defective samples. We then employ a DDPM equipped with a cross-attention mechanism to synthesize novel image-mask pairs, effectively expanding the dataset and improving downstream AD performance. III. SYSTEM ARCHITECTURE The proposed D-ADDA framework is organized into three stages (see Figure 2), each comprising multiple interacting subsystems that collaboratively overcome the limitations of AD in industrial production lines, particularly in cases where defective training samples are both limited and sensitive. We structure the stages as follows: •Stage 1: Data acquisition and augmentation system. This system collects data from the production lines on local nodes and performs data augmentation. •Stage 2: Generative modeling system. This system trains two generative models and distributes non-sensitive data. •Stage 3: Cross-conditional synthesis and detection system. This system performs defect synthesis on sensitive data and AD tasks. In the following, we provide a detailed description of each stage and its constituent subsystems. A. Stage 1: Data Acquisition and Augmentation System Deployed at each local production line, this system uses low-cost imaging devices to acquire samples consisting of normal and defective products. To compensate for limited defective samples, it includes a Data Augmentation Module (DAM) that operates on the collected data. DAM uses image processing and superpixel-based techniques to transfer defect structures onto normal images, generating a large set of synthetic defective samples. The system involves the following steps: •Local Data Acquisition. Each production node acquires Nimages using low-cost devices: Dlocal ={(Ii, Mi, yi)}N i=1,(1) where Ii∈RH×W×3is the image, Mi∈RH×W×1is the corresponding mask, and yi∈ {0,1}is the defect label. Hand Wdenote the image height and width, respectively. •Data Augmentation Module (DAM). We apply a superpixel-based function DAM(·)to enrich the defective sample space: Daug ={(I′ i, M′ i, yi)}N i=1,(2) where (I′ i, M′ i) = DAM(Ii, Mi). •Transmission of Non-Sensitive Data. Only synthetic samples without sensitive content (e.g., normal samples and masks) are transmitted to the central server: Dserver ={(I′ i, yi)|yi= 0}∪{(M′ i, yi)|yi= 1}.(3) 4 Fig. 2: Our D-ADDA framework integrates local augmentation, centralized diffusion modeling, and distributed model updates. Algorithm 1 Data Augmentation Module (DAM) Pipeline Require: n0: Number of samples to generate 1: procedure DATAAUGMENTATIONMODULE(DN, DD, DM) ▷DN: Normal samples, DD: Defective samples, DM: Defective masks 2: Randomly select a normal sample IN i, a defective sample ID jand a mask MD j. 3: SPi←EXTRACTSUPERPIXELS(IN i) 4: for k= 1 to n0do ▷Iterate for each new sample 5: IDaligned j←ALIGNIMAGES(IN i, ID j)▷Using cross-correlation 6: MDaligned j←ALIGNMASK(MD j, IDaligned j) 7: MDA j←REPLACESUPERPIXELS(MDaligned j, SPi) 8: MD∗ A j←AUGMENTMASK(MDA j) 9: MATCHPATCHHISTOGRAMS(IN i, IDaligned j, MD∗ A j) 10: ID∗ j←REPLACEPATCH(IN i, IDaligned j, MD∗ A j)▷ Generate novel defect 11: if QUALITYASSESSMENT(ID∗ j)then 12: Save ID∗ jand MD∗ A jin Daug 13: end if 14: end for return Daug ▷Return novel defective dataset 15: end procedure The DAM consists of a multi-step pipeline, outlined in Algorithm 1. The key stages are detailed below, where ID, MDand IN, MNdenote defective and normal images and masks, respectively: •Superpixel Extraction. Initially, the quick shift algorithm [39] is employed to extract superpixels from a randomly selected normal sample IN ibased on color space. These superpixels serve as potential patches for defect replacement. Quick Shift is a mode-seeking clustering approach in which each data point is linked to a local maximum or dense region in the underlying probability distribution. This method progressively moves each pixel toward its nearest neighbor with greater density, forming compact, consistent regions that conform to object contours and retain critical structural information. •Alignment Procedure. A cross-correlation-based alignment procedure is used to align a defective training image ID jwith the normal sample IN iusing a center cropped patch obtained from the defective image. •Mask Creation and Augmentation. Subsequently, the dot product between the defective mask MDand the superpixels is computed to identify defective superpixels which are then augmented by incorporating adjacent superpixels, thereby defining a larger defective area. To introduce variability, a random target superpixel and a specified percentage of the nearest superpixels p∼ U(p1, p2)are selected to form a new defective mask, which is then subjected to smoothing. Patches corresponding to this new mask are extracted from both normal and defective samples and normalized using histogram equalization. •Defective Sample Generation. Previous generated defective patch is superimposed onto the normal sample to create a new defective sample ID∗ j. •Data Quality Assessment (DQA). To assess the quality of the generated samples, we define a data quality assessment procedure. Specifically, we use a similarity function S(IA, IB)to determine if two images are similar. In our work, the Structural Similarity Index (SSIM) score is employed as follows: 1 m m X i=1 S(ID∗ j, ID i)≥1 n n X i=1 S(ID∗ j, IN i).(4) Here, mand ndenote the number of defective and normal samples, for a category (or class) C, respectively. In other words, if a novel defective image ID∗ jlies closer to the training defective images than to the normal images, we retain it along with its corresponding mask. Otherwise, we generate a new defective mask and repeat the process 5 GT (normal) GT (defect) GT (mask) −→ Superpixels Novel mask Extracted patch −→ Novel sample −→ −→ −→ −→ −→ −→ Fig. 3: Main steps of the local data augmentation pipeline for three categories: hazelnut (crack), cable (bent wire), and wood (scratch). GT stands for ground truth. Novel indicates newly generated synthetic samples that are distinct from the training set. for a fixed number of steps, after which we include the image in the augmented set. •Textures. Unlike objects, texture patches do not undergo the alignment phase, as almost each location achieves a maximum cross-correlation value without requiring alignment. For textures, we introduce additional randomness by randomly shifting the augmented mask in both the x and ydirections. A visualization of the main above steps and final results are shown in Figure 3. B. Stage 2: Generative Modeling System Housed on a central server, it is responsible for training generative diffusion models. It interacts with the local systems by receiving the augmented datasets without sensitive content generated in Stage 1. In this way, no sensitive information is exchanged between local and central nodes. In Stage 2, we train two generative models: an unconditional model GI trained on normal images INto capture the intrinsic characteristics of the category C, and a conditional model GMtrained on defective masks MDto learn plausible defect structures. To train GM, we use a classifier-free guidance to condition the architecture on specific defect classes [40]. The system involves the following steps: •Generative Model Training. The central server trains two generative models on the received non-sensitive data: GI(θI) : unconditional model on normal images IN; GM(θM) : conditional model on defective masks MD. θIand θMrepresent the parameters of the models. •Distribution of Non-Sensitive Weights. The learned model parameters, θIand θM, are transmitted back to the local nodes for downstream generation. More in detail, GIand GMrepresent two generative models, which aim to learn the true distribution of observed samples by representing uncertainty through latent variables. The models synthesize novel normal samples and masks, with GIused in a pre-training phase to capture domain characteristics and GMused to generate novel masks. In particular, we consider DDPMs as the architecture for generative models, since diffusion-based models have shown excellent performance and flexibility across various applications [10]. In our work, we let G= (E,D)represent a diffusion process [14] with an encoder-decoder structure following the U-Net architecture [36]. We use two separate models, GI= (EI,DI)and GM= (EM,DM), to generate normal images and masks, respectively. Differently than the standard U-Net, we connect the encoder and decoder via a Self-Attention (SA) module, equipped with multiple multi-head attention mechanisms to focus on specific areas of the data, allowing the models to learn diverse representations and enhance the quality and variety of the generated images. Specifically, given a batch of image features at layer Lwith height Hand width W, denoted as xl t∈RB×C×H×W, where Band Crepresent the batch size and number of channels, respectively, we define an attention function Attention(Xk t, Xq t, Xv t). Here, Xk t,Xq t, and Xv t correspond to the key, query, and value matrices used in our attention mechanism to compute attention at time step t. The input Xl tis projected into Qh t,Kh t, and Vh tfor each attention head h, with dimensionality dh, and passed to the attention function as follows: Qh t=Xl tWh Q, Kh t=Xl tWh K,(5) Ah t=softmax Qh t(Kh t)⊤ √d,(6) where Wh Qand Wh K∈Rdh×dfor h= 0,1, . . . , N −1, and d is the projected dimension of our attention mechanism. Then, attention coefficients Ah tare right multiplied by Vh t=Xl tWh V, where Wh V∈Rdh×d. We apply the SA module in both the intermediate and bottleneck layers of our U-Net while training GIand GM. After training, we obtain GIpre-trained on normal samples to capture the primary features of the category Cand GMtrained on augmented masks to generate novel defective masks. 6 C. Stage 3: Cross-Conditional Synthesis and Detection System This system is local and integrates the outputs of the two previously defined diffusion models to synthesize realistic defective images. Then, the generated images are used to train the classification and detection/localization model. Specifically, GIis modified to be conditioned on synthetic masks sampled from GM, using selfand cross-attention mechanisms to produce novel defective samples that reflect the distribution of real-world anomalies. The system involves the following steps: •Local Defect Synthesis. The local system fine tunes the generative model GI, pretrained in Stage 2, to synthesize novel defects employing selfand cross-attention mechanisms: z∼ N(0,I2), Mi=GM(z), Ii=GI(z|Mi),(7) •Defect Classification and Detection/Localization. We process the synthesized and acquired samples for downstream tasks: ˆyi=Fcls(Ii),ˆ Mi=Fdet/loc(Ii, Mi),(8) where Fcls and Fdet/loc denote the classification and localization/detection networks, respectively. Once the local nodes receive the weights θMand θIof both models trained on the central server, they first synthesize novel masks using the generative model GMand then fine-tune GIwith a Cross-Attention (CA) mechanism to condition the network on the defective masks. In this stage, we first process the input features xCA in , together with the conditioning mask MDand its complement M1−Drepresenting normal regions, through two separate convolutional-based mask embeddings ME(·). We then extract queries, keys, and values for MD and M1−D, respectively. We adopt a gating-based strategy to condition our model on input masks using two cross-attention layers, one self-attention layer and a residual connection as follows: GD=σ(GAP(Attention(Qt, KD t, V D t))), G1−D=σ(GAP(Attention(Qt, K1−D t, V 1−D t))), xSA =Attention(Qt, Kt, Vt).(9) Qt, Kt, Vtare obtained as projections of xCA in ,KD tand VD t are projected from ME(MD)while K1−D tand V1−D tfrom ME(M1−D). GAP denotes a global average pooling layer while σa sigmoid activation function. Finally, our output xCA out is computed as a linear combination of the original image features xCA in , the self-attention output xSA, and the gating terms GDand G1−D: xCA out =GD⊙(xSA +xCA in ) | {z } Defective Gate +G1−D⊙(xSA +xCA in ) | {z } Normal Gate , (10) where ⊙represents the element-wise multiplication. The proposed stage encourages the model to focus on both defective and non-defective regions while the gating mechanism separates their representations. A residual connection ensures better gradient flow. For added supervision, we project the defective mask into a similar space as the time projection and concatenate both embeddings into the U-Net’s downsampling and upsampling blocks. IV. EXPERIMENTAL RESULTS In the following, we present the details of the dataset used in our study, the model training parameters, and the evaluation protocols employed. We then describe the comparative methods and deployment scenarios considered, followed by quantitative and qualitative results for both classification and localization/detection tasks. Finally, we report on distributed experiments and ablation studies that examine the impact of key hyperparameters in the proposed framework. A. Experimental Setting Dataset. For our experiments, we use a well-established AD dataset, namely MVTec AD [41], which comprises 5354 high-resolution images across 15 categories (10 objects and 5 textures). We adhere to the standard evaluation protocol, which involves using one-third of the anomalous data as the training set, with the remainder reserved for testing. Additionally, we generate 500 samples per defect using our data augmentation module, with an image resolution of 128 ×128 pixels, and train the diffusion processes to produce 500 image/mask pairs per defect, resulting in a total of 36 500 image/mask pairs. Furthermore, we perform a quality check on randomly generated images over a specific number of iterations, i.e., 100. To prevent getting stuck in the quality inspection step, the last generated image is automatically added to the augmented dataset if none of the synthesized images pass the check. Model Training. We use 1000 diffusion steps and train the denoising U-Net for 30 000 steps with a learning rate of 10−4. As warm-up phase, we consider 5000 steps. Our mask embedding module comprises three Conv2D-ReLU-MaxPooling blocks, followed by an additional Conv2D layer to match the dimension of xCA in . We use 4heads for all the attention mechanisms and set dhto 32. Additionally, we apply a weighted exponential moving average with a weight decay of 0.995. The entire framework is trained using the Adam optimizer on a single NVIDIA RTX-6000 GPU. In our data augmentation module, we set p1and p2to 0.35/0.65 and 0.65/0.85 for textures/objects, respectively. Evaluation Protocol. To evaluate the quality of our synthesized images and masks, we focus on standard AD tasks: classification and detection/localization [7]. For classification, we consider each category in the MVTec AD dataset and perform a classification task among its defective classes. To assess the effectiveness of the generated images, we use a ResNet-based model to classify defective samples. Specifically, classification performance is measured as the proportion of correctly classified instances among all test instances across the Nddefect classes within a given category C. For detection and localization, the objective is to identify defective regions at both the image level and the pixel level. Detection focuses on a global assessment, determining whether an image is anomalous from the anomaly mask, while localization evaluates how well the predicted anomaly mask aligns with the actual defective regions across all pixels. A single U-Net architecture 7 TABLE I: CLASSIFICATION PERFORMANCE (%) Category U-ProF-Pro DiffAug CDC Crop&Paste SDGAN Defect-GAN DFMGAN DAM D-ADDA (NoP) D-ADDA DAM D-ADDA (NoP) D-ADDA bottle 48.84 38.76 52.71 48.84 53.49 56.59 72.09 79.07 81.40 67.44 65.12 83.72 cable 21.36 39.06 32.81 21.88 21.36 45.31 42.19 43.75 46.88 42.19 35.94 34.67 capsule 34.67 28.89 32.89 30.22 32.00 37.23 29.33 33.33 44.00 29.33 26.67 31.25 carpet 35.48 25.27 27.96 21.50 29.03 47.31 25.81 33.87 38.71 35.48 29.03 34.67 grid 28.33 35.83 28.33 30.83 27.50 40.83 20.00 27.50 32.50 12.50 17.50 25.00 hazelnut 65.28 54.86 59.03 43.75 61.11 81.94 62.50 95.83 97.92 64.58 95.83 93.75 leather 40.74 43.38 34.39 38.10 42.33 49.73 22.22 46.03 71.43 17.46 58.73 47.62 metal nut 58.85 48.44 59.89 44.27 56.77 64.58 76.56 67.19 73.44 59.38 64.06 65.62 pill 29.86 21.88 26.74 20.49 28.47 29.52 37.50 31.25 40.62 38.54 21.88 28.12 screw25.10 32.92 28.81 26.75 28.81 37.45 35.80 40.74 61.73 23.46 23.46 49.61 tile 59.65 48.54 68.42 42.69 26.90 74.85 64.91 85.96 94.74 75.44 75.44 75.44 transistor 38.09 29.76 41.67 32.14 35.72 52.38 57.14 60.71 64.29 50.00 50.00 57.14 wood 41.27 28.57 47.62 30.95 24.60 49.21 47.62 38.10 52.38 47.62 30.95 37.71 zipper 22.76 14.63 26.42 21.54 18.70 27.64 25.61 60.71 46.34 31.71 56.10 59.76 Average 39.31 35.06 40.55 32.43 34.77 49.61 44.23 53.15 60.46 42.51 46.48 51.72 Notes. U-Pro = Unfair protocol; F-Pro = Fair protocol; D-ADDA (noP) = we test our framework also w/o the ImageNet pre-training. is employed for both tasks, trained using a dataset comprising normal and defective samples. Performance in detection and localization is evaluated using three standard metrics: AUROC (Area Under the Receiver Operating Characteristic curve), AP (Average Precision), and F1-max. AUROC measures the trade-off between true positive rate and false positive rate across different thresholds. AP summarizes the precision-recall curve by computing a weighted average of precision at various recall levels. F1-max refers to the maximum F1score achieved across all possible thresholds. Since evaluations on the MVTec AD dataset are commonly performed using the test set [7], [35], we define two evaluation protocols to clarify the impact of tuning procedures. The first, named Unfair Protocol (U-Pro), uses the test set for fine-tuning model parameters, aligning with prior works. By contrast, we introduce a Fair Protocol (F-Pro), where performance is assessed using the model from the final training epoch, without test-time adaptation. We report results under these protocols for the classification task only, as detection and localization metrics showed limited sensitivity to the evaluation strategy. Methods. For a comprehensive evaluation, we compare our D-ADDA framework with DiffAug [42], CDC [43], Crop&Paste [44], SDGAN [45], Defect-GAN [6], and DFMGAN [5] on the classification task. For detection and localization, we evaluate D-ADDA against samples generated directly by our data augmentation module DAM, i.e., without employing any generative architecture. For the qualitative comparison, we consider AnomDiff [7] and DRÆM [35]. B. Results Classification Task. Table I reports the classification performance of our proposed framework under both unfair (UPro) and fair (F-Pro) evaluation protocols, including results without ImageNet pre-training (D-ADDA (NoP)). Under the U-Pro setting, D-ADDA achieves the highest accuracy in 10 out of 14 categories, reaching an overall average of 60.46%, which substantially outperforms both its unpretrained variant and all competing methods. The results highlight the strong contribution of ImageNet pre-training while also showing that the framework maintains competitive performance even without it. D-ADDA is also the best performing framework in particularly challenging categories such as hazelnut (97.92%), tile (94.74%), and bottle (81.40%), demonstrating robust generalization across diverse defect types and low-data regimes. Under the F-Pro protocol, which enforces fair evaluation conditions, D-ADDA again achieves the highest average accuracy, outperforming both DAM and D-ADDA (NoP). This indicates that our framework remains effective even when prior knowledge from large-scale datasets is removed, underscoring its ability to detect anomalies under realistic and unbiased evaluation settings. Additionally, D-ADDA maintains competitive results also in categories where prior methods had strong performance, such as metal nut, pill, and screw, reflecting the framework’s versatility and robustness across a wide spectrum of defect types. Detection/Localization Tasks. Table II presents a comparison between DAM and our D-ADDA framework on detection and localization AD tasks. Although it does not rely on any generative architecture, DAM outperforms most GAN-based methods and is therefore selected as the main competitor in this analysis. The evaluation metrics include AUC, AP, and F1-max across 15 object categories. For detection, D-ADDA outperforms DAM across all average metrics: AUC improves from 76.71% to 82.05%, AP from 86.87% to 90.35%, and F1-max from 83.59% to 84.93%. Additionally, significant improvements are observed in several categories. For example, in the wood category, AUC increases from 84.50% to 98.20%, AP from 93.40% to 99.30%, and F1-max from 86.00% to 96.30%, showing a substantial enhancement in AD. Similarly, the grid category sees AUC rise from 84.90% to 91.20%, and F1-max from 83.00% to 90.20%. The carpet category benefits from a jump in AUC from 66.60% to 77.40%, with AP also improving from 80.90% to 90.50%. In texturedominated categories like leather, carpet, and wood, D-ADDA consistently scores higher, confirming improved generalization across both structured and unstructured object types. Our framework also demonstrates clear advantages in accurately localizing defective regions. The average AUC increases from 79.83% to 81.63%, AP from 26.23% to 29.23%, and F1-max from 28.75% to 31.91%. Particularly large improvements are observed in categories with subtle or small defects. For instance, in the carpet category, F1-max improves from 8 TABLE II: DETECTION/LOCALIZATION PERFORMANCE (%) Detection Localization Category DAM D-ADDA DAM D-ADDA AUC AP F1-max AUC AP F1-max AUC AP F1-max AUC AP F1-max bottle 75.50 87.90 86.30 83.80 93.00 84.80 57.00 13.60 21.60 74.10 14.80 15.00 cable 70.20 76.20 68.80 69.70 77.60 69.20 88.20 30.90 35.50 72.70 26.50 39.40 capsule 66.60 85.40 87.20 59.20 81.70 86.70 89.40 06.00 09.60 91.90 06.30 09.80 carpet 66.60 80.90 81.90 77.40 90.50 82.10 68.60 13.00 18.10 88.50 28.20 30.20 grid 84.90 92.20 83.00 91.20 94.70 90.20 55.20 02.00 03.30 70.10 11.30 21.50 hazelnut 98.80 98.70 94.90 92.00 95.10 87.60 98.30 55.90 53.40 93.10 41.90 44.90 leather 81.30 91.20 82.80 85.80 93.40 85.30 58.50 04.90 10.80 77.50 19.30 26.40 metal nut 61.90 85.40 85.90 83.70 94.30 89.10 92.30 76.20 74.90 94.10 81.70 69.20 pill 70.70 91.20 88.50 76.60 92.80 89.60 87.90 58.20 58.30 88.60 58.10 60.00 screw80.00 90.80 82.40 80.20 90.60 83.40 92.30 06.40 08.50 77.20 01.70 04.40 tile 96.10 97.70 91.90 91.90 96.30 89.70 80.50 33.60 38.40 70.90 52.20 54.00 toothbrush 61.30 67.80 83.30 70.80 81.80 80.90 81.30 04.10 03.40 86.10 06.30 09.30 transistor 81.00 76.60 66.70 87.80 80.20 73.80 93.30 51.60 50.70 66.60 28.90 36.90 wood 84.50 93.40 86.00 98.20 99.30 96.30 70.90 22.50 22.10 85.40 46.00 46.90 zipper 71.20 87.70 84.20 82.50 93.90 85.20 83.80 14.60 22.70 87.60 15.20 10.70 Average 76.71 86.87 83.59 82.05 90.35 84.93 79.83 26.23 28.75 81.63 29.23 31.91 GT (defect) GT (mask) DRÆM AnomDiff D-ADDA Hazelnut (print) Leather (fold) Wood (combined) Fig. 4: Qualitative examples of localization performance for three categories. GT stands for ground-truth. 18.10% to 30.20%, and in leather, it increases from 10.80% to 26.40%. The grid category, which poses challenges due to its regular repeating patterns, sees F1-max increase from 3.30% to 21.50%, and AP from 2.00% to 11.30%. The wood category also demonstrates strong gains: F1-max improves from 22.10% to 46.90%, and AP from 22.50% to 46.00%. Even in low-performing cases for DAM such as toothbrush and screw, D-ADDA delivers consistent improvements across all three metrics. The results in Table II demonstrate that DADDA achieves robust improvements in both detection, which determines whether an object is anomalous, and localization, which identifies the specific anomalous regions. The gains are consistent across a wide variety of object geometries and textures, showing the method’s effectiveness in both coarse and fine-grained AD scenarios. Figure 4 provides qualitative examples for three categories considering centralized state-of-the-art generative methods. While DRÆM [35] excels at localizing the print defect on hazelnut objects, our approach competes closely with the pretrained AnomDiff [7] model, accurately identifying the central defect area. Localization on leather is more difficult, with all methods struggling to achieve high precision. For wood texture, our model appears to outperform DRÆM, although D-ADDA AnomDiff Hazelnut (crack) Capsule (crack) Grid (broken) Fig. 5: Qualitative comparison between synthesized images by D-ADDA (left) and AnomDiff (right). AnomDiff shows better results, but this comes at the cost of an extensive pre-training. Image Quality Inspection. Figure 5 shows a qualitative comparison between images generated by D-ADDA and a similar, yet centralized, diffusion-based one (i.e., AnomDiff [7]). The synthesized images generated through our augmentation process enhance the diffusion-based learning, leading to more realistic defective images and masks in specific cases. For example, the hazelnut object shows defects that appear highly realistic, whereas reproducing defects for the capsule object proves more challenging. Nevertheless, we observe some misalignment in texture-based objects, as shown in Figure 6, which arises from the absence of an explicit alignment loss within our architecture. While this misalignment benefits certain classes, e.g., wood or tile, due to the uniformity of their background, it negatively impacts localization metrics for grid patterns, where precision is more critical. Furthermore, certain generated images bear close resemblance to the training data, 9 Capsule (crack) Cable (broken) Grid (damaged case) Leather (cut) Fig. 6: Examples of misaligned image-mask pairs. raising concerns of overfitting; this issue could be alleviated by diversifying the augmented samples, for example through the addition of local nodes. Distributed Classification. To comprehensively evaluate the performance of our proposed method in distributed learning scenarios, we conduct a series of experiments inspired by the FedAvg algorithm [46]. Distributed learning provides a well-established framework for simulating collaborative training across multiple clients, making it a suitable and widely adopted proxy for distributed industrial inspection systems. While decentralized learning approaches offer benefits such as improved fault tolerance and reduced dependence on a central server, the lack of publicly available implementations and standardized benchmarks limits fair and reproducible comparisons. Therefore, we assume that model weights are averaged across clients when performing the classification task. As a baseline, we first consider the centralized scenario with a single client (NC= 1). In this setup, all data are available to one client, representing an idealized reference where no inter-client variability exists. This allows us to isolate the effects of distributed training by comparison. Then, to extend the classification to a distributed learning context, we simulate multiple clients operating within each production line. For example, in a scenario where Manufacturer A and Manufacturer B produce distinct products, each manufacturer may include multiple inspection stations acting as clients. To simulate multiple clients, we consider two client data partitioning strategies: •Non-IID (Label-wise Partitioning). Client datasets are partitioned by defect type, so that each client has access only to samples of a single class. Given Nddefect types in a category C, one client is assigned per defect. We optimize local models on class-specific data and aggregate their updates using the averaging algorithm. This setup simulates real-world inspection stations specialized for specific defect classes, resulting in highly skewed data distributions. •IID (Uniform Partitioning). The entire dataset is randomly shuffled and evenly divided across NCclients, ensuring each client receives a representative sample of all classes. This serves as a baseline scenario, representing idealized homogeneous data distributions among clients. TABLE III: DISTRIBUTED CLASSIFICATION ACCURACY Distributed Parameters Accuracy [%] No Pre-training Pre-training Num. Clients B E IID U-ProF-ProU-ProF-Pro 1– – – 53.15 46.48 60.46 51.72 Nd81✗28.10 18.21 26.12 21.04 Nd16 1✗24.80 18.32 26.69 22.60 Nd32 1✗25.76 21.79 26.84 20.17 Nd85✗25.30 20.78 24.61 21.52 Nd16 5✗25.68 20.94 24.38 22.11 Nd32 5✗24.94 18.54 23.81 20.91 Nd810 ✗27.49 21.35 23.84 20.25 Nd16 10 ✗27.08 22.10 24.81 22.34 Nd32 10 ✗25.47 20.45 24.33 20.12 Nd81✓31.66 28.48 45.32 43.04 Nd16 1✓27.64 24.74 40.24 38.61 Nd32 1✓26.56 23.10 36.11 34.68 Nd85✓37.62 33.27 48.20 44.68 Nd16 5✓35.63 30.46 48.08 44.80 Nd32 5✓32.38 29.72 42.72 40.63 Nd810 ✓39.79 35.30 49.99 47.21 Nd16 10 ✓35.94 31.59 47.00 44.23 Nd32 10 ✓33.54 27.83 44.32 42.22 10 81✓27.43 25.13 41.07 38.84 10 16 1✓27.46 23.85 39.39 37.42 10 32 1✓25.70 21.31 31.73 28.28 10 85✓35.90 31.48 46.82 44.06 10 16 5✓32.41 28.79 47.62 45.27 10 32 5✓28.26 25.03 41.90 39.29 10 810 ✓38.24 33.22 45.69 43.55 10 16 10 ✓36.78 30.68 46.10 43.70 10 32 10 ✓31.31 27.85 45.35 42.92 Notes. B= local batch size; E= local number of epochs. For the non-IID configuration, the number of clients is set to Nd, corresponding to one client per defect class. For the IID configuration, experiments were conducted with NC=Ndand NC= 10. We systematically vary the number of local training epochs (E∈ {1,5,10}) and batch sizes (B∈ {8,16,32}). Table III reports the classification accuracy. In the centralized scenario (NC= 1), the model has access to all data, yielding a reference point for performance without interclient heterogeneity. In the non-IID regime, where each client receives samples from only a single defect class, accuracy remains consistently low (20–27%), indicating that this scenario struggles to reconcile highly divergent local updates. Adjustments of Bor Eproduce minimal improvement: smaller batches fail to introduce sufficient gradient diversity, and additional local epochs increase divergence between client models. In the IID regime, where each client receives a balanced subset of all defect classes, accuracy improves substantially. For NC=Nd,B= 8, and E= 10,U-Pro reaches 49.99% and F-Pro reaches 47.21%. Performance generally benefits from more local epochs, particularly for smaller batch sizes, reflecting improved convergence when inter-client gradient variance is low. Larger batch sizes remain competitive but occasionally result in slightly lower accuracy due to reduced stochasticity. When NCis fixed at 10, accuracy is slightly lower than the NC=Ndcase, with the difference more pronounced at lower epoch counts. The qualitative visualizations in Figure 7 for the hazelnut and tile categories confirm the quantitative trends observed