Turn-Anchored Robustness Tokens for Multi-Turn Multimodal Safety

Vision-language assistants now hold long conversations in which users can attach images at any turn. Two failure modes compound in that setting. First, safety guardrails erode over turns: a request that would be refused outright can be sliced into innocent precursor steps, and by turn eight the model complies — SafeMT reports GPT-4o's attack success rate rising from 15.39% in a single turn to over 50% in an eight-turn conversation. Second, the same long, mixed-modality trajectories are an ideal hiding place for backdoors: POISONSHARE showed that a text-only trigger can be split across $k$ of $n$ turns so that every individual turn looks benign, and input-aware dynamic triggers $T(x)$ Input-Aware Dynamic Backdoor Attack defeat every classical backdoor defense precisely because no two poisoned samples share a footprint. Nobody has yet combined these: there is no attack in the literature whose trigger is simultaneously input-aware, distributed across turns, and split across modalities — and therefore no defense has ever been tested against one.

Defending by full adversarial retraining is not an option. The min-max inner loop that makes adversarial training work on CIFAR-scale models becomes exponential in dialogue depth when the adversary may perturb any turn's image or text, and the one multi-turn alignment method that works, multi-round red-teaming RL MTSA, is by its authors' own account computationally intensive. This proposal builds the missing threat and a cheap defense against it in one program: a handful of learnable tokens, tuned on a completely frozen model in roughly a thousand optimization steps, that anchor each turn's internal safety representation against single-turn corruption and cancel composite trigger activations at the final turn.

Core Claim

The claim is that safety against both semantic multi-turn drift and distributed dynamic backdoors can be bought for the cost of tuning about ten token vectors, not the cost of retraining a foundation model. Concretely, we extend Robustness Tokens — which restored a frozen Vision Transformer's adversarial robustness with 10 tokens in 200 steps — from static single images to conversation state, and we add a contrastive term borrowed from Decayed Contrastive Decoding (DCD), which neutralized POISONSHARE's text backdoor by contrasting the final layer against an intermediate "amateur" layer, reducing attack success on Mistral-7B from 89% to 0.35%. Our hypothesis: turn-level feature anchoring handles drift and single-turn perturbations, while the layer-contrast term catches the composite activations that only fire when a full $k$-of-$n$ trigger assembles at the final turn — all without searching the exponential space of turn combinations.

The Threat: POISONSHARE-MM

We first construct the evaluation vehicle, because no benchmark in the literature contains it. Following POISONSHARE's protocol (poisoning a fraction of multi-turn instruction-tuning dialogues) but generalizing the trigger along two axes from the dynamic-backdoor literature, we poison a 7–8B open vision-language model during instruction tuning. Each planted trigger is generated per-dialogue by a small conditioned generator, in the spirit of the autoencoder trigger network of Input-Aware Dynamic Backdoor Attack, and its components are split across turns and modalities: an imperceptible image perturbation on turn 1, a typographic fragment FigStep on turn 3, an innocuous lexical marker on turn 5. The poisoned behavior activates only when the composite state assembles. We will verify that the text-only POISONSHARE result (>99% attack success at 5% poisoning) carries over to the multimodal setting, that mismatched or partial triggers are ignored (the cross-trigger discipline of Input-Aware Dynamic Backdoor Attack reported 88–97% of mismatched triggers classified as clean), and that STRIP-style entropy checks, fine-pruning, and turn-level moderation filters all fail — the last because every turn is individually safe, exactly the property that makes the SIST (safe image + safe text) cases of SIUO and USB so hard for external inspection.

The Defense: Turn-Anchored Robustness Tokens

The defender appends $m \approx 10$ learnable token vectors to the model's input at every turn and tunes only those vectors, with the entire backbone frozen. Let $\phi$ denote the token parameters, $s_t$ the dialogue state (all text and images through turn $t$), $z_\phi(s)$ the model's pooled safety-relevant state representation, $s^{\mathrm{adv}}_t$ the same dialogue with exactly one turn's image or text replaced by a one-step worst-case perturbation, $p_\phi(\cdot \mid s)$ the final-layer next-token distribution, and $p^{\mathrm{mid}}_\phi(\cdot \mid s)$ the distribution read out from an intermediate amateur layer selected by Jensen–Shannon divergence as in DCD. Training minimizes

$$ \min_{\phi}\; \sum_{t=1}^{n} \big\| z_\phi(s^{\mathrm{adv}}_t) - z_0(s^{\mathrm{clean}}_t) \big\|_2^2 \;+\; \lambda\, \mathrm{D}_{\mathrm{JS}}\big( p_\phi(\cdot \mid s^{\mathrm{adv}}_n),\, p^{\mathrm{mid}}_\phi(\cdot \mid s^{\mathrm{adv}}_n) \big), $$

over a few thousand multi-turn multimodal dialogues taken from the SafeMT and MTMCS-Bench pools, with $\lambda$ fixed by a small sweep. The first term anchors every turn's state so that corrupting one turn cannot move the safety representation — the mechanism that let ROB tokens recover white-box classification accuracy on DINOv2-G from 0.3% to 63.1% while holding clean accuracy at 85.2% versus 85.6%. The second term suppresses final-layer activations that the amateur layer does not share, the signature DCD exploits against distributed triggers. The key move on cost: instead of an inner-loop adversary over full trajectories (exponential in $n$), we use one-shot single-turn perturbations as a surrogate and let the contrast term catch what the surrogate misses. This is a relaxation of the true min-max and we treat it as a hypothesis, not a guarantee; the in-context amortization argument of Adversarially Pretrained Transformers May Be Universally Robust In-Context Learners is our best reason to expect tokens trained on cheap surrogates to transfer to composite threats they never saw.

Baselines

The do-nothing reference point is the unmodified VLM evaluated on SafeMT (10,000 samples, 17 harmful scenarios), where even GPT-4o degrades from 15.39% single-turn attack success to over 50% by turn eight; our defense must at minimum beat that curve on the same dialogues. The strongest external competitor is SafeMT's own Dialogue Safety Moderator, an auxiliary Gemma-3-4B-it model fine-tuned to inject guard prompts mid-conversation; we will run it on identical SafeMT and MTMCS-Bench dialogues (12,032 dialogues built from 752 COCO images) against the same target model, with an equal tuning-data and GPU-hour budget. Beating it means lower multi-turn attack success and Safety Index at strictly lower added latency — our tokens add one short sequence per turn, not a 4B forward pass — without increased over-refusal on XSTest. The strongest training-based competitor is Adversary-Aware DPO (ADPO), which reached a 99.5% safe response rate under white-box attack on MMPGDBlank versus 67% for standard DPO and 24% for SFT; it is single-turn, so we will extend it to multi-turn dialogues ourselves, retrained on our same data with the same hyperparameter search budget, and count our method as winning only if it matches ADPO's attack success within an order of magnitude less tuned parameters and optimization steps. Against the backdoor threat specifically, the strongest competitor is DCD from POISONSHARE (89% to 0.35% on text), ported to the VLM's decoding. Finally, the key ablation isolates our ingredient: tokens trained with the anchoring loss on single-turn states only, versus full turn-anchored training with the contrast term — if the single-turn variant performs as well, the multi-turn machinery is unjustified and we will say so.

Why not just…?

Why not just put a guardrail classifier on every turn? This is the most serious alternative, and it partially works: external moderation is exactly SafeMT's answer. But MTMCS-Bench evaluated five existing guardrails, including AdaShield, and found they "mitigate some failures but do not fully resolve multi-turn contextual risks" while hurting helpfulness; and the SIST results of USB and SIUO (where the best model, GPT-4V, reached only a 53.26% safe rate) show that individually safe inputs compose into hazards an inspector never sees. POISONSHARE-MM is deliberately constructed so every turn passes turn-level inspection. The experiments will make this concrete by running the strongest moderator on our poisoned dialogues and reporting its miss rate.

Why not just write a stronger system prompt? Safety prompting is free and works in single-turn settings — MM-SafetyBench reports cutting LLaVA-1.5-7B's attack success from 77.33% to 15.68% on its tiny benchmark. But attention to a system prompt dilutes as history grows Jailbreaking in the Haystack, roleplay strategies overwrite instructions outright, and ADPO shows prompt- and SFT-level defenses reach only a 24% safe rate under white-box perturbations. We include safety prompting as a zero-cost baseline and will report honestly any regime where it suffices.

Why not just fine-tune with standard DPO on multi-turn safe dialogues? Standard DPO treats the visual embedding space and each turn as static, achieving 67% versus ADPO's 99.5% under white-box attack, and MTSA showed turn-by-turn optimization fails against multi-round attacks — only future-reward RL cut Zephyr-7B from 74.50% to 23.50%. Aggressive alignment also buys over-refusal: USB found the safest model (Claude 3.5 Sonnet, 15.57% ASR) was also the most oversensitive at 25.82% refusal on harmless prompts. The real threat here is the compute-matched ADPO-on-multi-turn baseline above; if it closes most of the gap at comparable cost, our contribution reduces to the threat model, and the experiments are designed to surface exactly that outcome rather than argue around it.

Why not just run classical backdoor defenses? Trigger inversion, entropy analysis, and pruning are all invalidated by input-aware dynamic triggers Input-Aware Dynamic Backdoor Attack, and any defense that searches turn combinations faces the exponential $k$-of-$n$ bottleneck POISONSHARE was built around. We will still run STRIP-style checks as sanity baselines and expect them to fail; that is part of the threat demonstration, not the defense.

Proposed Experiments

Experiment 1 establishes the threat: poison an open 7–8B VLM with POISONSHARE-MM at 5% poisoning, and report attack success, partial-trigger rejection, and evasion of turn-level filters, comparing against POISONSHARE's text-only >99% at the same poisoning rate; no reported multimodal analogue exists on Lacuna, which is the point. Experiment 2 tests semantic multi-turn safety on SafeMT and MTMCS-Bench: multi-turn attack success and Safety Index versus the moderator and ADPO baselines, over-refusal on XSTest (targeting an increase below MTSA's 5.62%), and helpfulness on MT-Bench. Experiment 3 tests generalization to threats excluded from token training: the typography attacks of MM-SafetyBench (where LLaVA-1.5-7B jumps from 5.25% to 79.38% on illegal activity), the SIST cases of USB (many open models above 80% ASR), and SIUO's 269 cases. Experiment 4 addresses the honest open question from BadPrompt, which achieved over 97% attack success by poisoning just 2 continuous-prompt samples: can an adversary plant a backdoor in our tokens themselves, and does the defense survive if the token values are public, the known secrecy weakness of ROB tokens? Experiment 5 is the compute accounting that makes or breaks the claim: tuned parameters, optimization steps (target: order 10 tokens and order 1,000 steps against ROB's 200-step precedent), and GPU-hours versus the ADPO and moderator baselines at matched data. The decisive check is a single plot: multi-turn attack success versus tuning cost, where our method must sit at or below the best baseline's error at a tenth of its tuning cost — or the core claim is falsified.