Adaptive Gradient Shielding for Consistent Defense Against Adversarial Attacks in Deep Learning

Deep learning models are highly vulnerable to adversarial perturbations, posing serious risks in critical domains such as healthcare, autonomous systems, and cybersecurity. This has motivated the need for robust defenses that maintain reliability without significant computational overhead. Current defense strategies fail to provide a unified solution for gradient-based and gradient-free adversarial attacks. Many approaches suffer from gradient masking or accuracy degradation during clean inference. Methods that can adaptively stabilize model responses under diverse and evolving adversarial conditions have a significant defense gap. This study introduces AGS, a lightweight defense that combines perturbation-aware feature smoothing with dynamic gradient modulation. Using an adaptive auxiliary controller, AGS detects locally sensitive features and selectively regulates gradient flow, improving robust generalization without full adversarial training. The proposed method is evaluated on CIFAR-10 and ImageNet-Subset under FGSM, PGD-20, and BIM attacks. AGS significantly improves robustness while preserving clean accuracy. On CIFAR-10, the PGD-20 accuracy increased from 42.8% to 71.4% with only a 0.4% drop in clean accuracy, and the FGSM accuracy improved from 61.5% to 84.7%. On the ImageNet-Subset, the PGD-20 accuracy increased from 18.6% to 39.3%. The auxiliary controller adds only 3.2% computational overhead, which is lower than that of standard adversarial training, demonstrating that AGS is an efficient and adaptive defense for varying adversarial conditions.

Adaptive shielding; Adversarial robustness; Deep learning security; Gradient modulation; Perturbation resilience

Abasi, A. K., Aloqaily, M., & Guizani, M. (2025). 6G mmWave security advancements through federated learning and differential privacy. IEEE Transactions on Network and Service Management, 22(2), 1911–1928. https://doi.org/10.1109/TNSM.2025.3528235

Abbes, A., Slimani, N., Cherif, C., Hadjout, S., Mammasse, A., Mankouri, A., & Kherbachi, S. (2025). Adversarial attacks and defense mechanisms in computer vision: A comprehensive survey. Communications in Computer and Information Science, 2462, 371–378. https://doi.org/10.1007/978-3-031-88226-5_26

Alajaji, A. (2025). FortinIDS: Defending smart city IoT infrastructures against transferable adversarial poisoning in machine learning-based intrusion detection systems. Sensors, 25(19), 6056. https://doi.org/10.3390/s25196056

Al-Andoli, M., Tan, S., Sim, K., Goh, P., & Lim, C. (2024). A framework for robust deep learning models against adversarial attacks based on a protection layer approach. IEEE Access, 12, 17522–17540. https://doi.org/10.1109/ACCESS.2024.3354699

Anjum, S., & Nnaji, C. (2025). Enhancing robustness of deep learning models against adversarial attacks for construction worker safety. Automation in Construction, 179, 106447. https://doi.org/10.1016/j.autcon.2025.106447

Chelliah, B., Malik, M., Kumar, A., Singh, N., & Regin, R. (2023). Similarity-based optimised and adaptive adversarial attack on image classification using neural network. International Journal of Intelligent Engineering Informatics, 11(1), 71. https://doi.org/10.1504/IJIEI.2023.130715

Costa, J., Roxo, T., Proença, H., & Morais Inácio, P. (2024). How deep learning sees the world: A survey on adversarial attacks & defenses. IEEE Access, 12, 61113–61136. https://doi.org/10.1109/ACCESS.2024.3395118

Dadhwal, H., de Abreu, M., Parvizi, N., & Saha, S. (2026). Benchmarking the adversarial resilience of machine learning models for DDoS detection. Array, 29, 100664. https://doi.org/10.1016/j.array.2025.100664

Elabd, E. (2025). Dynamic differential privacy technique for deep learning models. Scientific Reports, 15(1), 39353. https://doi.org/10.1038/s41598-025-27708-0

Guo, Z., Qian, Y., Zhao, S., Dong, J., Li, Y., Arandjelovi?, O., Fang, L., & Lau, C. (2026). Artwork protection against unauthorized neural style transfer and aesthetic color distance metric. Pattern Recognition, 171, 112105. https://doi.org/10.1016/j.patcog.2025.112105

He, K., Kim, D., & Asghar, M. (2023). Adversarial machine learning for network intrusion detection systems: A comprehensive survey. IEEE Communications Surveys & Tutorials, 25(1), 538–566. https://doi.org/10.1109/COMST.2022.3233793

Hu, J., Wang, Z., Shen, Y., Lin, B., Sun, P., Pang, X., Liu, J., & Ren, K. (2024). Shield against gradient leakage attacks: Adaptive privacy-preserving federated learning. IEEE/ACM Transactions on Networking, 32(2), 1407–1422. https://doi.org/10.1109/TNET.2023.3317870

Jhajharia, K., & Shrivastavaa, Y. (2024). Adversarial attacks and defenses on deep learning models. 2024 3rd International Conference for Advancement in Technology (ICONAT), 1–5. https://doi.org/10.1109/ICONAT61936.2024.10775002

Jiang, L., Ma, L., & Yang, G. (2025). Shadow defense against gradient inversion attack in federated learning. Medical Image Analysis, 105, 103673. https://doi.org/10.1016/j.media.2025.103673

Kim, Y., Jung, J., Kim, H., So, H., Ko, Y., Shrivastava, A., Lee, K., & Hwang, U. (2024). Adversarial defense on harmony: Reverse attack for robust AI models against adversarial attacks. IEEE Access, 12, 176485–176497. https://doi.org/10.1109/ACCESS.2024.3505215

Laila, D. A. (2025). Responsive machine learning framework and lightweight utensil of prevention of evasion attacks in the IoT-based IDS. STAP Journal of Security Risk Management, 2025(1), 59–70. https://doi.org/10.63180/jsrm.thestap.2025.1.3

Li, S., Wang, J., Wu, H., Zhang, J., Cheng, X., Luo, X., & Ma, B. (2025). Defense against adversarial faces at the source: Strengthened faces based on hidden disturbances. IEEE Transactions on Artificial Intelligence, 6(7), 1761–1775. https://doi.org/10.1109/TAI.2025.3527923

Manzoor, A., Fargetta, G., Ortis, A., & Battiato, S. (2026). ShieldNet: A novel adversarially resilient convolutional neural network for robust image classification. Applied Sciences, 16(3), 1254.

Muoka, G., Yi, D., Ukwuoma, C., Mutale, A., Ejiyi, C., Mzee, A., Gyarteng, E., Alqahtani, A., et al. (2023). A comprehensive review and analysis of deep learning-based medical image adversarial attack and defense. Mathematics, 11(20), 4272. https://doi.org/10.3390/math11204272

Ness, S. (2025). Enhancing remaining useful life prediction against adversarial attacks: An active learning approach. IEEE Access, 13, 170921–170934. https://doi.org/10.1109/ACCESS.2025.3611453

Puttagunta, M., Ravi, S., & Nelson Kennedy Babu, C. (2023). Adversarial examples: Attacks and defences on medical deep learning systems. Multimedia Tools and Applications, 82(22), 33773–33809. https://doi.org/10.1007/s11042-023-14702-9

Rahman, M., Roy, P., Frizell, S., & Qian, L. (2025). Evaluating pretrained deep learning models for image classification against individual and ensemble adversarial attacks. IEEE Access, 13, 35230–35242. https://doi.org/10.1109/ACCESS.2025.3544107

Roy, A., & Dasgupta, D. (2025). Defending learning systems against mean-shift perturbations. IEEE Transactions on Artificial Intelligence, 1–13. https://doi.org/10.1109/TAI.2024.3422929

Saha, S., Das, S., & Carvalho, G. (2026). A hybrid deep learning model for adversarially resilient internet traffic prediction. Computers and Electrical Engineering, 130, 110832. https://doi.org/10.1016/j.compeleceng.2025.110832

Shan, F., Lu, Y., Li, S., Mao, S., Li, Y., & Wang, X. (2025). Efficient adaptive defense scheme for differential privacy in federated learning. Journal of Information Security and Applications, 89, 103992. https://doi.org/10.1016/j.jisa.2025.103992

Shayea, G., Zabil, M., Habeeb, M., Khaleel, Y., & Albahri, A. (2025). Strategies for protection against adversarial attacks in AI models: An in-depth review. Journal of Intelligent Systems, 34(1). https://doi.org/10.1515/jisys-2024-0277

Wang, C., Liu, Y., Liu, X., He, Y., & Xiao, K. (2025). Dynamic shielding: Defense mechanism against gradient attacks on distributed GANs. SSRN. https://doi.org/10.2139/ssrn.5251836

Wang, F., Hugh, E., & Li, B. (2024). More than enough is too much: Adaptive defenses against gradient leakage in production federated learning. IEEE/ACM Transactions on Networking, 32(4), 3061–3075. https://doi.org/10.1109/TNET.2024.3377655

Wang, Y., Sun, T., Li, S., Yuan, X., Ni, W., Hossain, E., & Vincent Poor, H. (2023). Adversarial attacks and defenses in machine learning-empowered communication systems and networks: A contemporary survey. IEEE Communications Surveys & Tutorials, 25(4), 2245–2298. https://doi.org/10.1109/COMST.2023.3319492

Wei, W., & Liu, L. (2022). Gradient leakage attack resilient deep learning. IEEE Transactions on Information Forensics and Security, 17, 303–316. https://doi.org/10.1109/TIFS.2021.3139777

Xiangfei, Z., & Qingchen, Z. (2025). Defending against attacks in deep learning with differential privacy: A survey. Artificial Intelligence Review, 58(11), 347. https://doi.org/10.1007/s10462-025-11350-3

Xu, F., Wang, C., Liang, J., Zuo, C., Yue, K., & Li, W. (2024). A knowledge distillation strategy for enhancing the adversarial robustness of lightweight automatic modulation classification models. IET Communications, 18(14), 827–845. https://doi.org/10.1049/cmu2.12793

Yinusa, A., & Faezipour, M. (2025). A multi-layered defense against adversarial attacks in brain tumor classification using ensemble adversarial training and feature squeezing. Scientific Reports, 15(1), 16804. https://doi.org/10.1038/s41598-025-00890-x

Zhang, C., & Wang, P. (2026). Adaptive masked autoencoder defense: A robust defense strategy against adversarial attacks in deep learning models. Journal of Circuits, Systems and Computers, 35(1). https://doi.org/10.1142/S0218126625503578

Zhao, W., Alwidian, S., & Mahmoud, Q. (2022). Adversarial training methods for deep learning: A systematic review. Algorithms, 15(8), 283. https://doi.org/10.3390/a15080283

Zhou, S., Liu, C., Ye, D., Zhu, T., Zhou, W., & Yu, P. (2023). Adversarial attacks and defenses in deep learning: From a perspective of cybersecurity. ACM Computing Surveys, 55(8), 1–39. https://doi.org/10.1145/3547330