ProjLens: Unveiling the Role of Projectors in Multimodal Model Safety

Kun Wang, Cheng Qian, Miao Yu, Lilan Peng, Liang Lin, Jiaming Zhang, Tianyu Zhang, Yu Cheng, Yang WangarXiv 2026 · arXiv

TL;DR

ProjLens is a novel interpretability framework that opens the black box of backdoor attacks in Multimodal Large Language Models (MLLMs), focusing specifically on the projector layer. The authors discover that backdoor mechanisms are hidden within a strictly low-rank subspace of the projector's weights and operate via a "universal drift vector" in the embedding space, whose activation magnitude scales linearly with the input feature norm.

What Problem Does This Paper Address?

Multimodal Large Language Models (MLLMs) integrate vision encoders with LLMs using a connecting module called a projector. Because projectors are parameter-efficient, they are the primary target for fine-tuning and downstream task alignment. However, this architecture introduces a dangerous attack surface: data poisoning backdoor attacks. An adversary can inject malicious behaviors (like jailbreaks or forced refusals) that activate only when a specific visual trigger is present, leaving normal capabilities intact.

While we know these attacks are highly effective, the underlying mechanics have remained opaque. Prior interpretability research mostly focused on model capabilities (like attention analysis) rather than safety, leaving a critical gap: How exactly do these backdoors store and execute their malicious instructions within the projector?

Methods

The authors introduce ProjLens, an interpretability framework that dissects the lifecycle of multimodal backdoors across three dimensions:

  1. Feature Space (Visual Trigger Probe): The authors propose a learnable Visual Trigger Probe (VTP)—a lightweight 3-layer MLP binary classifier—to determine if trigger-induced features are cleanly separated from benign features within the projector's output embeddings.
  2. Weight Space (SVD & Neuron Analysis): They apply Singular Value Decomposition (SVD) to the weight updates ($\Delta W$) before and after backdoor injection to analyze its mathematical rank. They also track individual neuron activation frequencies and magnitudes to see if specific "trigger neurons" exist.
  3. Embedding Space (Dynamic Transformations): They analyze the projected residual ($\Delta E$) between pre- and post-poisoned embeddings. Using LogitLens, they decode these embedding shifts directly into the LLM's vocabulary space to read the semantic changes.

Datasets

The framework evaluates four distinct backdoor configurations across standard datasets:

  • Targeted Refusal: Evaluated on VQAv2, using global Gaussian noise as a trigger to force an AI refusal response.
  • Malicious Injection: Evaluated on Flickr30k, using a local green patch trigger to append a malicious suffix to captions.
  • Perceptual Hijack: Evaluated on MSCOCO Caption, using a specific icon trigger to force the model to describe an entirely different target image.
  • Jailbreak Output: Evaluated on VLBreakBench, using an oil painting style transfer trigger to bypass safety alignments.

To ensure general model utility wasn't destroyed by the backdoor, they benchmarked performance using MathVista (mathematical reasoning) and POPE (object hallucination robustness).

Key Insights and Ideas

  • The Projector Weight Paradox: Globally, the weight updates from backdoor training look completely normal, diffuse, and full-rank, lacking any obvious "trigger neurons." However, the backdoor functionality is actually tightly packed into a low-rank subspace.
  • The Trojan Projection Hypothesis: Backdoors do not create a completely separate pathway for triggered images. Instead, both clean and poisoned samples undergo a semantic shift toward the exact same malicious direction (a universal drift vector).
  • Norm-Driven Activation: If clean images also drift toward the backdoor target, why don't they trigger it? The authors found that the drift magnitude scales linearly with the $L_2$ norm of the input tokens. Visual triggers artificially inflate local feature norms, pushing poisoned samples past the threshold into full backdoor activation, while clean samples remain unaffected during standard greedy sampling.

Key Results

  • High Attack Efficacy: Restricting backdoor fine-tuning solely to the projector yields massive Attack Success Rates (ASR), reaching 98.30% for Malicious Injection and 97.00% for Perceptual Hijack, while actually improving accuracy on clean downstream tasks.
  • Latent Risks Discovered: Even when clean samples show a 0% explicit ASR, their latent probability of generating the backdoor target ($P_{bkd}$) spikes significantly (e.g., jumping from 13.30% to 50.87% in Targeted Refusal), revealing a hidden bias.
  • Defensive Mitigation: Because the backdoor is low-rank, removing just the rank-1 component of the weight residual dramatically mitigates the attack—restoring model utility in the Jailbreak task from 28.91% to 86.91%. Conversely, injecting a mere rank-3 approximation of the backdoor weights into a clean model successfully reconstructs the attack with up to 90.7% ASR.
[Clean/Poisoned Input] -> [Projector Layer] -> [Universal Drift Vector v0]
                                                     |
                                        Linear Scaling via L2 Norm
                                                     |
                                 -------------------------------------
                                 |                                   |
                       (Low Norm: Clean)                    (High Norm: Triggered)
                                 |                                   |
                       [Normal Generation]                  [Backdoor Activated]

Limitations and Future Work

The primary limitation highlighted is the dual-use risk of the research: while these insights provide a blueprint for low-rank backdoor defenses, malicious actors could theoretically exploit them to design stealthier, low-rank attacks explicitly optimized to evade SVD-based detection.

For future work, the authors aim to leverage the ProjLens framework to develop robust, training-time defenses and precise, interpretability-based alignment strategies to permanently patch these vulnerabilities.

Summary updated

Abstract

Multimodal Large Language Models (MLLMs) have achieved remarkable success in cross-modal understanding and generation, yet their deployment is threatened by critical safety vulnerabilities. While prior works have demonstrated the feasibility of backdoors in MLLMs via fine-tuning data poisoning to manipulate inference, the underlying mechanisms of backdoor attacks remain opaque, complicating the understanding and mitigation. To bridge this gap, we propose ProjLens, an interpretability framework designed to demystify MLLMs backdoors. We first establish that normal downstream task alignment--even when restricted to projector fine--tuning--introduces vulnerability to backdoor injection, whose activation mechanism is different from that observed in text-only LLMs. Through extensive experiments across four backdoor variants, we uncover:(1) Low-Rank Structure: Backdoor injection updates appear overall full-rank and lack dedicated ``trigger neurons'', but the backdoor-critical parameters are encoded within a low-rank subspace of the projector;(2) Activation Mechanism: Both clean and poisoned embedding undergoes a semantic shift toward a shared direction aligned with the backdoor target, but the shifting magnitude scales linearly with the input norm, resulting in the distinct backdoor activation on poisoned samples. Our code is available at: https://anonymous.4open.science/r/ProjLens-8FD7
BibTeX
@misc{wang2026projlens,
  author = {Kun Wang and Cheng Qian and Miao Yu and Lilan Peng and Liang Lin and Jiaming Zhang and Tianyu Zhang and Yu Cheng and Yang Wang},
  title = {{ProjLens: Unveiling the Role of Projectors in Multimodal Model Safety}},
  year = {2026},
  eprint = {2604.19083},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.19083},
}

Loading related content...

Figure