REVIEW 14 cited by
Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Current vision large language models (VLLMs) exhibit remarkable capabilities yet are prone to generate harmful content and are vulnerable to even the simplest jailbreaking attacks. Our initial analysis finds that this is due to the presence of harmful data during vision-language instruction fine-tuning, and that VLLM fine-tuning can cause forgetting of safety alignment previously learned by the underpinning LLM. To address this issue, we first curate a vision-language safe instruction-following dataset VLGuard covering various harmful categories. Our experiments demonstrate that integrating this dataset into standard vision-language fine-tuning or utilizing it for post-hoc fine-tuning effectively safety aligns VLLMs. This alignment is achieved with minimal impact on, or even enhancement of, the models' helpfulness. The versatility of our safety fine-tuning dataset makes it a valuable resource for safety-testing existing VLLMs, training new models or safeguarding pre-trained VLLMs. Empirical results demonstrate that fine-tuned VLLMs effectively reject unsafe instructions and substantially reduce the success rates of several black-box adversarial attacks, which approach zero in many cases. The code and dataset are available at https://github.com/ys-zong/VLGuard.
Forward citations
Cited by 14 Pith papers
-
Gradient Immunity: Null-Space Resistance to Malicious Fine-Tuning
Calibrating a null-space gate to fully cover the defender's harmful data keeps post-fine-tuning attack success at pre-release levels, but this is a coverage and calibration consequence rather than a tested defense on ...
-
One Anchor for All: Unified Multilingual and Multimodal Safety Alignment for LVLMs
A small set of model neurons shared across languages and modalities can transfer English-only safety training to multilingual and multimodal refusal behavior.
-
V-DEAL: Diagnosing Video Safety De-Calibration as an Understanding-Refusal Coupling Failure
Video LLMs understand harmful content but activate weaker refusal signals when the query is benign; prompt realignment reduces attack success from ~48% to ~1%.
-
Safe responses matter: Output-aware safety guardrail mitigate over-refusal in MLLMs
A lightweight multi-instance contrastive classifier on MLLM hidden states predicts output harmfulness, matching input-side safety while sharply lowering over-refusal.
-
Secure Tug-of-War (SecTOW): Iterative Defense-Attack Training with Reinforcement Learning for Multimodal Model Security
An iterative attacker-defender reinforcement learning method that makes a multimodal LLM refuse more jailbreak prompts without over-refusing ordinary queries.
-
LoX: Low-Rank Extrapolation Robustifies LLM Safety Against Fine-tuning
Low-rank extrapolation of an aligned model's weight update (LoX) reduces how much later fine-tuning erodes safety refusal behavior.
-
Bootstrapping LLM Robustness for VLM Safety via Reducing the Pretraining Modality Gap
Reducing the modality gap between image and text embeddings during LVLM pretraining reduces unsafe response rates by up to 16.3% across models and benchmarks.
-
MM-RLHF: The Next Step Forward in Multimodal LLM Alignment
A human-annotated multimodal preference dataset plus critique-based reward modeling and reward-margin-weighted DPO improves MLLM performance across many benchmarks.
-
Attention Misses Visual Risk: Risk-Adaptive Steering for Multimodal Safety Alignment
Risk-adaptive activation steering, guided by a prototype-similarity risk score computed on the first three response tokens, substantially reduces multimodal jailbreak success rates across four MLLMs while preserving utility.
-
Seeing the Threat: Vulnerabilities in Vision-Language Models to Adversarial Attack
A two-stage evaluation framework and token-projection analysis show that LVLMs encode harmful semantic cues from images even without OCR, while remaining vulnerable to cross-modal attacks.
-
VSCBench: Bridging the Gap in Vision-Language Model Safety Calibration
A new benchmark, VSCBench, measures oversafety and undersafety in vision-language models and shows that most models, including proprietary ones, are miscalibrated on at least one safety dimension.
-
A Multimodal Automatic Redteaming Evaluation based on Atomic Jailbreak Strategy Decoupling and Combination
A new jailbreak framework, HACA, combines atomic text and image attack strategies selected by a cross-modal planner and generates attacks with LLMs and text-to-image models, reaching 95.48% average attack success acro...
-
A Survey of Safety on Large Vision-Language Models: Attacks, Defenses and Evaluations
A survey of LVLM safety that adds a lifecycle taxonomy and new benchmark results showing Janus-Pro-7B has weaker safety than several open-source LVLMs.
-
When Data Manipulation Meets Attack Goals: An In-depth Survey of Attacks for VLMs
A survey that classifies VLM attacks by goal and data manipulation strategy, and reviews defenses and metrics.
Discussion (0). Continue with ORCID to comment.