While autoencoders are pivotal in critical applications such as anomaly detection and medical imaging, their reliability remains understudied compared to supervised models. Although mutation testing has advanced for neural networks, no framework exists for assessing autoencoder robustness against real-world faults, leaving a gap in safety-critical validation. Furthermore, autoencoders lack explicit labels, rendering traditional mutation metrics ineffective. We propose MuAE, the first mutation testing tool tailored for autoencoders, addressing this gap through: (1) a fault taxonomy derived from real-world debugging cases; (2) mutation operators that inject faults while preserving model validity; and (3) reconstruction error (Erec) as an evaluation metric to quantify fault impacts on output fidelity. We validate MuAE on CIFAR-10 using two convolutional autoencoders and four mutation operators (M1, M2, M3), generating 3 mutants per operator. Mutations significantly degraded performance: layer reinitialization increased Erecby up to 210%, while weight noise caused milder degradation. This condensed evaluation demonstrates the feasibility of mutation-based robustness assessment for unsupervised models and highlights potential links to adversarial vulnerability.

Adversarial training (AT) is widely regarded as a leading defense strategy for improving the robustness of deep learning models against adversarial attacks. However, existing AT methods often rely on a single attack strategy during training, which limits the exploration of the perturbation space and leads to poor generalization robustness against stronger, unseen, or adaptive adversarial attacks. Moreover, most AT approaches overlook class-wise robustness–the observed variation in robustness across different image classes–by focusing solely on average performance over the entire dataset. In this paper, we present Advanced Distributional Training with Class-wise Robustness (ADT++), a novel adversarial training framework that significantly improves generalization robustness against unseen and sophisticated adversarial attacks. Following the standard adversarial training framework, ADT++ is formulated as a minmax optimization problem, where the inner maximization aims to learn the worst-case adversarial distribution around adversarial examples to further explore the perturbation space. The outer minimization seeks to find model parameters that minimize the expected loss of the maximum inner loss. To further improve the generalization robustness, ADT++ leverages the class-wise robustness phenomenon by targeting the most vulnerable image classes with high-loss adversarial attacks to generate more impactful adversarial examples. Extensive evaluations on benchmark datasets and against various AT defense methods and adversarial attacks confirm the effectiveness of ADT++ in improving model robustness against stronger and adaptive attacks. The source code of ADT++ can be found.11https://github.com/LAiSR-SK/ADT2Plus

Rapid detection of bacterial contamination in food supply chains is essential to prevent outbreaks and protect public health. Traditional microbiological methods, such as culture and biochemical testing, are slow and can delay necessary interventions. This study presents deep learning–based approaches for the early detection and classification of Escherichia coli and Salmonella Typhimurium in microscopy images captured during incubation periods ranging from 1.5 hours to 4 hours, with a step size of 30 min, under two different conditions: a plain background and an onion mixture background. We annotated a dataset of 2,200 high-resolution (2160×1620) images, collected under 60× magnification from the time frame 1.5 h to 4 h, using a combination of manual and semi-automated techniques with Mask R-CNN. For manual annotation, dataset was annotated using tkinter framework in python. For model training, only pure culture data were used. The initial pipeline employed Cellpose for segmentation and a ViT for classification across bacterial growth stages. However, Cellpose was unable to segment all instances in an image due to its requirement for a diameter parameter, which, under variable colony sizes, resulted in partial segmentation. To overcome these challenges, we adopted an end-to-end Mask R-CNN model for instance segmentation. Mask R-CNN achieved consistently strong mean Intersection over Union (mIoU) scores between 0.91 and 0.98 across growth stages. Over the full incubation period, the mean Average Precision (AP) and Average Recall (AR) at IoU threshold 0.5 were 0.93 and 0.96 for E. coli, and 0.945 and 0.975 for Salmonella, respectively, indicating robust detection performance across bacterial types and bacterial growth period. Using fine-tuned Mask R-CNN trained on background aware time-stamped microscopy datasets, our method achieves an [email protected] of 0.95 for colonies cultured at 2 hours timepoint. This work focuses on proactive food safety monitoring in agricultural pipelines.