Improved Adversarial Training Stability in Complex Visual Tasks
Keywords:
adversarial training, visual learning, robust accuracy, gradient penalty, spectral normalization, dynamic loss weighting, training stabilityAbstract
Adversarial training is often used to improve the robustness of visual models. However, it may cause unstable gradients, large loss changes, and model collapse in complex visual tasks. This study presents a stable adversarial training method using dynamic loss weighting, spectral normalization, and gradient penalty. The dynamic loss weight controls the role of clean and adversarial samples during training. Spectral normalization limits model sensitivity, and gradient penalty reduces excessive parameter updates. Experiments were conducted on CIFAR-10 and an ImageNet subset under FGSM and PGD attacks. The perturbation strengths were set to 8/255 and 16/255. Compared with standard adversarial training, the proposed method improved robust accuracy by 6.5%–11.2%, reduced training loss fluctuation by about 34%, and improved FID by about 7.8%. These results show that the method can make adversarial training more stable while keeping useful visual features. This study provides a practical training strategy for robust visual learning in safety-related tasks, such as medical imaging, autonomous driving, and industrial inspection.
References
1. Chen, F., Li, S., Liang, H., Xu, P., & Yue, L. (2025). Optimization Study of Thermal Management of Domestic SiC Power Semiconductor Based on Improved Genetic Algorithm.
2. Kothuri, S. R., & Thaoroijam, K. (2026). A Robust Deep Ensemble Architecture for Safety‐Critical Decision‐Making in Autonomous Systems. Transactions on Emerging Telecommunications Technologies, 37(2), e70333.
3. Yan, B. (2026). A Comparative Study: Greedy versus GAN-Based Compressed Sensing for Fast MRI Reconstruction. International Journal of Advance in Applied Science Research, 5(3), 36-42.
4. Javed, H., El-Sappagh, S., & Abuhmed, T. (2024). Robustness in deep learning models for medical diagnostics: security and adversarial challenges towards robust AI applications. Artificial Intelligence Review, 58(1), 12.
5. Liang, R., Feifan, F. N. U., Liang, Y., & Ye, Z. (2025). Emotion-Aware Interface Adaptation in Mobile Applications Based on Color Psychology and Multimodal User State Recognition. Frontiers in Artificial Intelligence Research, 2(1), 51-57.
6. Mohan, A., & Schön, T. (2026). Towards Robust Agents: A Survey of Adversarial Attacks and Defenses in Deep Reinforcement Learning. IEEE Access.
7. Zhu, W., Yao, Y., & Yang, J. (2025). Optimizing Financial Risk Control for Multinational Projects: A Joint Framework Based on CVaR-Robust Optimization and Panel Quantile Regression.
8. Gordienko, Y., Nowakowski, G., Kochura, Y., Taran, V., & Stirenko, S. (2025, October). Generative Data Augmentation by Dataset Distillation. In International Workshop on Artificial Intelligence for Biomedical Data (pp. 105-118). Cham: Springer Nature Switzerland.
9. Feng, H. (2024, October). Design of Intelligent Charging System for Portable Electronic Devices Based on Internet of Things (IoT). In 2024 5th International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (ICBAIE) (pp. 568-571). IEEE.
10. Holla, H., Polepalli, S. R., & Sasikumar, A. A. (2025). Adversarial threats to cloud IDS: robust defense with adversarial training and feature selection. IEEE Access.
11. Liu, S., Liu, X., & Feng, H. (2025, November). Research on AI-Driven Visual Design and Immersive Interactive Experiences Based on Multimodal Cognition and User. In Proceedings of the 2025 International Conference on Digital Society and Intelligent Computing (pp. 734-740).
12. Madabhushi, S., & Dewri, R. (2025). Mitigating Over-Generalization in Anomalous Power Consumption Detection using Adversarial Training. ACM Transactions on Cyber-Physical Systems, 9(3), 1-25.
13. Chen, X., Xiao, H., Zeng, Z., Zhang, S., & Xiao, R. (2025). Fine-Grained Multivariate Time Series Anomaly Detection via Causal Inference. Knowledge-Based Systems, 114765.
14. Kose, K., & Zhou, B. (2025). Adversarial Training for Aerial Disaster Recognition: A Curriculum-Based Defense Against PGD Attacks. Electronics, 14(16), 3210.
15. Gao, X., & Lan, Y. (2025, December). Data-Driven Identification Method for Low-Voltage Distribution Station-User Relationships Based on Voltage Sequence Correlation. In 2025 7th International Academic Exchange Conference on Science and Technology Innovation (IAECST) (pp. 508-512). IEEE.
16. Gaur, S., Singh, A., Kansal, V., & Prakash, S. (2025, December). MART+ AWP: A Hybrid Adversarial Training Framework to Reduce Robust Overfitting. In 2025 International Conference on Decision Aid Sciences and Applications (DASA) (pp. 1684-1689). IEEE.
17. Gao, X., Chen, J., & Huang, M. (2026, January). Research on Risk Dependency Structures and Resource Allocation Optimization in New Energy Technology Collaboration within Enterprise Distributed Innovation. In 2026 International Conference on Electrical Engineering, Intelligent Control and Artificial Intelligence (EEICAI) (pp. 43-47). IEEE.
18. Hassija, V., Palanisamy, B., Chatterjee, A., Mandal, A., Chakraborty, D., Pandey, A., ... & Kumar, D. (2025). Transformers for vision: A survey on innovative methods for computer vision. Ieee Access.
19. Xu, S., Ma, Q., Liu, H., & Yue, L. (2026). Continuous Reorganization and Performance Preservation of Agent Memory Structure Under Distributed Change Environments.
20. Singh, R. K. (2025). Generative Adversarial Networks for Data Augmentation in Low-Resource NLP Tasks.
21. Liu, S., & Yim, J. (2025). Research on Generative AI Creation Systems Based on Visual Language Modeling: Human-Machine Collaboration and Cognitive Feedback Mechanisms. Available at SSRN 6139770.
22. Shreeya, K. N., Subburaj, B., Saketh, K. S. G., Padmavathy, T. V., Alphonse, S., & Subramanian, G. (2026). A quantum resilient deepfake detection framework using enhanced resnext and post quantum cryptography defence. Scientific Reports.
23. Gao, G., Gao, R., Gao, R., & Zhou, H. (2026). Engineering Analysis and Quantitative Research on the Platform-Based Evolution of Enterprise Communication Systems.
24. Dwivedi, P., Islam, B., & Kajal, M. (2025). Smooth gradient loss: a loss function for gradient regularization in deep learning optimization. The Journal of Supercomputing, 81(15), 1-45.
25. Wang, Y., Yin, X., Chen, J., & Wang, Y. (2026). Evidence-Based Study on Low-Burden Digital Phenotyping for Precision Screening of Oral Anti-Obesity Drug Efficacy.
26. Gu, X. (2026, January). Analysis of the Synergistic Role of User Natural Language Feedback and Behavioral Data in Adjusting Recommendation Strategies for Mental Health Platforms. In Proceedings of the 2026 5th International Conference on Big Data, Information and Computer Network (pp. 198-203).
27. Jiao, Y., Wang, A., Zhao, B., & Shi, T. (2026). The Impact of Visual Language Strategies in Public Art Creation on Community Spatial Perception and Public Behavior.
28. Haref, Q. M., Long, J., & Yang, Z. (2025). Enhanced Video-Based Fall Detection Using 3D Human Pose Estimation and Advanced Generative Adversarial Networks: Integrating WGAN Data Augmentation with a Hybrid 3D CNN–GRU Architecture. KSII Transactions on Internet & Information Systems, 19(12).
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Emily Carter, Thomas Whitfield, Charlotte Bennett
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
