Multi-modal large language models (MLLMs) have made significant progress, yet their safety alignment remains limited. Typically, current open-source MLLMs rely on the alignment inherited from their language module to avoid harmful generations. However, the lack of safety measures specifically designed for multi-modal inputs creates an alignment gap, leaving MLLMs vulnerable to vision-domain attacks such as typographic manipulation. Current methods utilize a carefully designed safety dataset to enhance model defense capability, while the specific knowledge or patterns acquired from the high-quality dataset remain unclear. Through comparison experiments, we find that the alignment gap primarily arises from data distribution biases, while image content, response quality, or the contrastive behavior of the dataset makes little contribution to boosting multi-modal safety. To further investigate this and identify the key factors in improving MLLM safety, we propose finetuning MLLMs on a small set of benign instruct-following data with responses replaced by simple, clear rejection sentences. Experiments show that, without the need for labor-intensive collection of high-quality malicious data, model safety can still be significantly improved, as long as a specific fraction of rejection data exists in the finetuning set, indicating the security alignment is not lost but rather obscured during multi-modal pretraining or instruction finetuning. Simply correcting the underlying data bias could narrow the safety gap in the vision domain. Warning: This paper contains harmful images and AI-generated contents which may be offensive.

Deep neural networks have played a crucial part in many critical domains, such as autonomous driving, face recognition, and medical diagnosis. However, deep neural networks are facing security threats from backdoor attacks and can be manipulated into attacker-decided behaviors by the backdoor attacker. To defend the backdoor, prior research has focused on using clean data to remove backdoor attacks before model deployment. In this paper, we investigate the possibility of defending against backdoor attacks by utilizing test-time partially poisoned data to remove the backdoor from the model. To address the problem, a two-stage method TTBD is proposed. In the first stage, we propose a backdoor sample detection method DDP to identify poisoned samples from a batch of mixed, partially poisoned samples. Once the poisoned samples are detected, we employ Shapley estimation to calculate the contribution of each neuron's significance in the network, locate the poisoned neurons, and prune them to remove backdoor in the models. Our experiments demonstrate that TTBD removes the backdoor successfully with only a batch of partially poisoned data across different model architectures and datasets against different types of backdoor attacks.

Deep neural networks have achieved impressive performance in a variety of tasks over the last decade, such as autonomous driving, face recognition, and medical diagnosis. However, prior works show that deep neural networks are easily manipulated into specific, attacker-decided behaviors in the inference stage by backdoor attacks which inject malicious small hidden triggers into model training, raising serious security threats. To determine the triggered neurons and protect against backdoor attacks, we exploit Shapley value and develop a new approach called Shapley Pruning (ShapPruning) that successfully mitigates backdoor attacks from models in a data-insufficient situation (1 image per class or even free of data). Considering the interaction between neurons, ShapPruning identifies the few infected neurons (under 1 % of all neurons) and manages to protect the model's structure and accuracy after pruning as many infected neurons as possible. To accelerate ShapPruning, we further propose discarding threshold and ∊ -greedy strategy to accelerate Shapley estimation, making it possible to repair poisoned models with only several minutes. Experiments demonstrate the effectiveness and robustness of our method against various attacks and tasks compared to existing methods.

An off-the-shelf model as a commercial service could be stolen by model stealing attacks, posing great threats to the rights of the model owner. Model fingerprinting aims to verify whether a suspect model is stolen from the victim model, which gains more and more attention nowadays. Previous methods always leverage the transferable adversarial examples as the model fingerprint, which is sensitive to adversarial defense or transfer learning scenarios. To address this issue, we consider the pairwise relationship between samples instead and propose a novel yet simple model stealing detection method based on SAmple Correlation (SAC). Specifically, we present SAC-w that selects wrongly classified normal samples as model inputs and calculates the mean correlation among their model outputs. To reduce the training time, we further develop SAC-m that selects CutMix Augmented samples as model inputs, without the need for training the surrogate models or generating adversarial examples. Extensive results validate that SAC successfully defends against various model stealing attacks, even including adversarial training or transfer learning, and detects the stolen models with the best performance in terms of AUC across different datasets and model architectures. The codes are available at https://github.com/guanjiyang/SAC.