Adversarial Training (AT) is widely acknowledged as one of the most effective strategies for defending against adversarial attacks. However, current AT methods often fail to significantly enhance the robustness of Deep Neural Network models and tend to compromise natural accuracy due to the intermixing of clean and adversarial examples. To overcome these limitations, we propose an Adaptive adversarial training method based on Mahalanobis Distance (AMD). Firstly, the method introduces an adversarial example generation strategy that adapts attack parameters to create a unique strategy for each example. This approach fine-tunes the attack intensity, thereby enhancing robustness accuracy. Secondly, a Mahalanobis distance constraint is introduced to prevent the overlap of clean and adversarial examples, with the attack strategy adjusted accordingly to preserve natural accuracy. The experimental results demonstrate that AMD effectively defends against five common attacks across three widely used computer vision datasets, outperforming six existing defense methods and significantly enhancing both robustness accuracy and natural accuracy. Specifically, on the CIFAR-10 dataset using WideResNet, the robust accuracy achieves 81.17% in the Projected Gradient Descent evaluation, while natural accuracy reaches 95.33%.

Large language models (LLMs), with their powerful natural language processing capabilities, can provide more advanced intelligent services for edge devices. However, deploying LLMs at the edge is vulnerable to jailbreak attacks, which can cause the model to generate unsafe content. Meanwhile, current jailbreak attack schemes are inefficient in generating highly stealthy jailbreak prompts. To address this, we propose a Fast and Controllable Bias-Guided Jailbreak Attack (FCB) scheme. First, to improve attack efficiency, we optimize the bias of the model’s output layer to guide the model in generating low-energy jailbreak prompts by directly adjusting the output layer’s logits, thereby accelerating the decoding process. Second, to enhance the stealthiness of the generated jailbreak prompts, we design token stop selection and bias normalization methods to constrain the perturbations during the iterative process, preventing the generation of jailbreak prompts without meaningful semantics. Finally, extensive experimental results demonstrate that FCB can generate highly stealthy jailbreak prompts within a short time. Specifically, compared to the current state-of-the-art controllable attack generation scheme, COLD Attack, FCB achieves up to a 8% improvement in attack success rate, reduces perplexity by up to 181.171, and shortens generation time by as much as 28 seconds.