Neural networks, despite their remarkable performance in widespread applications, including image classification, are also known to be vulnerable to subtle adversarial noise. Although some diffusion-based purification methods have been proposed, for example, DiffPure, those methods are time-consuming. In this paper, we propose One Step Control Purification (OSCP), a diffusion-based purification model that can purify the adversarial image in one Neural Function Evaluation (NFE) in diffusion models. We use Latent Consistency Model (LCM) and ControlNet for our one-step purification. OSCP is computationally friendly and time efficient compared to other diffusion-based purification methods; we achieve defense success rate of 74.19\% on ImageNet, only requiring 0.1s for each purification. Moreover, there is a fundamental incongruence between consistency distillation and adversarial perturbation. To address this ontological dissonance, we propose Gaussian Adversarial Noise Distillation (GAND), a novel consistency distillation framework that facilitates a more nuanced reconciliation of the latent space dynamics, effectively bridging the natural and adversarial manifolds. Our experiments show that the GAND does not need a Full Fine Tune (FFT); PEFT, e.g., LoRA is sufficient.

We propose Instruct2Attack (I2A), a language-guided semantic attack that generates semantically meaningful perturbations according to free-form language instructions. We make use of state-of-the-art latent diffusion models, where we adversarially guide the reverse diffusion process to search for an adversarial latent code conditioned on the input image and text instruction. Compared to existing noise-based and semantic attacks, I2A generates more natural and diverse adversarial examples while providing better controllability and interpretability. We further automate the attack process with GPT-4 to generate diverse image-specific text instructions. We show that I2A can successfully break state-of-the-art deep neural networks even under strong adversarial defenses, and demonstrate great transferability among a variety of network architectures.

Recent studies have shown that robustness to adversarial attacks can be transferred across networks. In other words, we can make a weak model more robust with the help of a strong teacher model. We ask if instead of learning from a static teacher, can models "learn together" and "teach each other" to achieve better robustness? In this paper, we study how interactions among models affect robustness via knowledge distillation. We propose mutual adversarial training (MAT), in which multiple models are trained together and share the knowledge of adversarial examples to achieve improved robustness. MAT allows robust models to explore a larger space of adversarial samples, and find more robust feature spaces and decision boundaries. Through extensive experiments on CIFAR-10 and CIFAR-100, we demonstrate that MAT can effectively improve model robustness and outperform state-of-the-art methods under white-box attacks, bringing $\sim$8% accuracy gain to vanilla adversarial training (AT) under PGD-100 attacks. In addition, we show that MAT can also mitigate the robustness trade-off among different perturbation types, bringing as much as 13.1% accuracy gain to AT baselines against the union of $l_\infty$, $l_2$ and $l_1$ attacks. These results show the superiority of the proposed method and demonstrate that collaborative learning is an effective strategy for designing robust models.