When Redaction Makes Vision-Language Models Guess

Before a photo reaches a vision-language model (a system that reads an image and writes text about it), privacy tools often black out, blur, or cover faces, documents, screens and license plates. These tools are judged on two numbers: how much private content they hide, and how well the model still answers questions about the rest of the image. This proposal argues that the two-number picture misses the most important effect. A redacted region is exactly the situation in which these models are known to fall back on language habits and describe what is usually there instead of what they see. When the model fills a covered region, a wrong fill is a hallucination and a right fill is a privacy leak recovered by inference. Both come from the same behavior: completing the gap from statistical priors.

We propose to measure this "redaction-induced completion" directly, to separate its two outcomes, and to test whether the style used to render the redaction, together with a lightweight inference-time correction, can suppress completion without costing utility on the visible content. The practical result would be a three-way trade-off curve covering privacy, utility and faithfulness for visual sanitization. We would also get evidence on which redaction styles invite guessing and which do not.

Core Claim

Hiding pixels does not hide information from a model with strong language priors. It changes an extraction problem into a guessing problem. We hypothesize four things. First, completion rate rises with mask area and with how strongly the masked object co-occurs with its visible surroundings. Second, the rise shows up in faithfulness metrics well before it shows up in utility metrics. Third, generic anti-hallucination methods reduce wrong completions but leave correct completions, which are leaks, largely intact. Fourth, rendering the redaction as a labeled placeholder, instead of a bare black box, blur or inpainting, gives the model something grounded to say about the gap and reduces both outcomes.

The motivating evidence comes from both directions. On the hallucination side, Evaluating Object Hallucination in Large Vision-Language Models showed that object hallucinations follow training co-occurrence statistics and worsen under adversarial co-occurrence polling. Villa et al. documented "hidden hallucinations," in which models answer correctly from context without genuine recognition. That is precisely the mechanism by which a covered object could be inferred.

On the privacy side, Connecting Pixels to Privacy and Utility built segmentation-based redaction before modern vision-language models existed. It found that privacy jumps like a step function once enough pixels are covered, while utility declines gradually, so over-redaction looks cheap. We claim that this conclusion depends on the absence of a generative reader. With one present, larger masks give more room to guess, and the cost appears as fabrication rather than as lost answers.

Beyond Blanket Masking is the closest modern work. It masks only high-risk fields on private objects and reports gains in model answerability and content identification over full-object masking. It measures whether the model can answer, not whether what it says about the masked parts is true. We keep its setting but change the outcome being measured.

Research Program

Controlled measurement on COCO. The first component is a controlled benchmark built from MSCOCO, where segmentation masks give full ground truth. We redact chosen objects, including non-private ones, so that the mechanism can be studied with exact labels. For every caption we then sort mentioned objects into three bins. An object is grounded if it is visible. It is a completion if it is present only under the mask, which counts as recovery and plays the role of leakage. It is a hallucination if it is absent from the image entirely.

This extends the instance-level caption metric of Object Hallucination in Image Captioning with the middle bin. We also build a three-answer version of POPE polling ("Is there a [object]?") for masked objects, where the correct answer is "cannot tell." Negatives are drawn with POPE's random, popular and adversarial schemes. We prefer polling for headline claims because POPE's prompt-template standard deviation (0.78) is much lower than the caption metric's (3.22) in the same paper.

Transfer to real private content. The second component moves to real privacy categories. We use the Visual Redactions Dataset, which provides pixel masks over 24 privacy attributes on VISPR images. We also use BIV-Priv-Seg, which contains private objects photographed by blind and low-vision users. Here leakage is scored by whether the output states an annotated private attribute of the masked region, for example the apparent age or gender behind a blurred face, a document type, or text content.

Because automatic vision-language judges themselves hallucinate errors that are not there (Chen et al., Kaiyi Huang), leakage and fabrication labels on the private sets will be assigned by human annotators on a stratified sample. Automatic scoring will be used only where segmentation ground truth exists.

Redaction style and mask size. The third component varies the redaction rendering and the mask area. The rendering styles are black box, Gaussian blur, pixelation, generative inpainting, and a labeled placeholder. The placeholder is a flat patch with rendered text of the category, such as "redacted: document." It is the natural-image analogue of the type-preserving tokens that Anonymization-Enhanced Privacy Protection for Mobile GUI Agents substitutes into screen hierarchies. That work relied on structured UI text and never studied free-form description of natural scenes.

We sweep mask dilation from tight segmentation to full bounding box and beyond. This lets us trace completion as a function of covered area and of co-occurrence strength, which we compute from COCO object statistics.

Inference-time correction. The fourth component is a correction for cases where the rendering cannot be changed. Following SHARP, which finds directions in middle-layer activations (layers 10–15) that separate faithful from hallucinated states and adds a vector along them at inference time, we extract one "complete-versus-abstain" direction. Activations are collected on COCO redaction pairs: outputs that describe the masked region versus outputs that acknowledge it is hidden. We then apply the direction unchanged to VISPR and BIV-Priv-Seg images.

Because the vector is mined only from public COCO data, no private image influences it. This sidesteps the concern raised by Differentially Private Steering for Large Language Model Alignment, which showed that text-model steering vectors derived from sensitive examples can memorize them. As a secondary check, we build an in-domain vector from VISPR using that paper's per-example clipping and Gaussian noise. This tells us whether private in-domain data buys enough accuracy to justify the privacy cost.

Models and compute. All experiments use LLaVA-1.5 and LLaVA-Next, the models on which the cited mitigation numbers were reported, plus LLaMA 3.2-11B Vision, which was used for privacy detection in Assessing Visual Privacy Risks in Multimodal AI. Every method gets the same images, masks, prompts and decoding settings. Methods with hyperparameters, meaning the steering strength and RITUAL's transform set, get the same tuning budget on a held-out COCO split.

Baselines

Naive reference points. The first reference is the unredacted image, which gives an upper bound on utility and the model's native hallucination rate. The second is the redacted image with plain black-box full-object masking and no mitigation. This second point is the default deployed practice, and every other method is measured against it.

Fine-grained masking. On BIV-Priv-Seg, the fine-grained masking of Beyond Blanket Masking is the strongest privacy-side baseline. It reports up to 11% higher answerability and up to 45% better content identification than full-object masking. We rerun its masks when they are available, and otherwise approximate them by masking only the high-risk sub-regions annotated in the dataset. Beating it means lower completion and hallucination at equal or better answerability on visible content. We expect partial masks may actually invite more completion, because more context survives around the gap.

Hallucination-side methods. The strongest hallucination-side baselines are two. RITUAL (RITUAL) decodes over the original image and a randomly transformed copy. It reports POPE accuracy of 88.87% versus 84.13% for base LLaVA-1.5 on unredacted images. SHARP's generic hallucination vector reports a sentence-level caption hallucination rate reduced from 52.8 to 34.8 on LLaVA-v1.5, also on unredacted images. There is no reported result found on Lacuna for any hallucination-mitigation method evaluated on redacted or masked images, so both will be run by us under identical conditions.

The key comparison is on the three-bin metric. Our method must match these methods on hallucinated completions and clearly beat them on correct completions, meaning leaks. It must also stay within a small margin of them on visible-object POPE accuracy.

Prompt disclosure. The prompt-disclosure baseline appends an instruction stating that dark or labeled patches are redacted and must not be described.

Ablations. Two ablations isolate the ingredients. The first applies the placeholder rendering alone, without steering, and the steering alone on black-box masks. The second replaces the complete-versus-abstain vector with SHARP's generic vector. This tests whether redaction needs its own direction or whether generic prior suppression suffices.

Why not just…?

Why not just tell the model in the prompt that masked regions are off limits? This is the real threat. It costs nothing, and redaction is at least explicitly visible, unlike the implicit norms tested elsewhere. Evidence against its sufficiency comes from PrivacyLens. There, models with near-perfect probing accuracy (GPT-4 at 98.99%) still leaked in 25.68% of actions, and privacy-enhancing prompts left leakage above 25%. Assessing Visual Privacy Risks in Multimodal AI found that taxonomy prompting helped one model family and hurt another.

Neither result concerns redacted images, so we treat prompting as a first-class baseline rather than dismissing it. If prompt disclosure matches placeholder-plus-steering on completion and leakage across all three models, we will report that the cheap fix suffices. In that case the contribution reduces to the measurement protocol.

Why not just inpaint the redacted region so the model sees a plausible, harmless scene? Inpainting removes the gap by inserting content that is not in the photo. By construction it converts potential leaks into guaranteed fabrications that downstream users cannot distinguish from real content. It is included as a rendering condition, and we predict it scores well on the utility and leakage axes and worst on the hallucination bin. That outcome would show exactly why a two-axis evaluation is misleading.

Why not just apply an existing anti-hallucination method such as RITUAL or SHARP? These methods restore attention to visual evidence. When the evidence is deliberately removed, restoring attention to the rest of the scene may strengthen the contextual cues from which correct completions are inferred. RITUAL's flips and crops also leave the mask in place. They are run as baselines, and the three-bin metric decides the question directly. If they suppress leaks as well as hallucinations, the separate complete-versus-abstain direction is unnecessary and we will say so.

Why not just fine-tune the model to respect redactions? Assessing Visual Privacy Risks in Multimodal AI reports that LoRA fine-tuning on privacy data produced reasoning hallucinations, with models inventing violations and attributes to justify their flags. More RLHF, More Trust? found that preference alignment raised privacy leakage by 12% and cut TruthfulQA accuracy by 25%. Fine-tuning is therefore a documented source of the very failure we measure, so it is not a clean fix. We will still include a small LoRA baseline on the COCO redaction pairs used to build the steering vector, with matched data, to test whether training on the same signal does better or worse than steering.

Proposed Experiments and Decisive Checks

Does redaction cause completion? On masked COCO objects, three-answer POPE should show that models assert the presence of masked objects well above the rate at which they assert truly absent random objects. Adversarially co-occurring masked objects should be completed most often. If completion rate does not rise with co-occurrence strength, the prior-completion account is wrong.

Does over-redaction stay cheap? We plot visible-content utility, via POPE on unmasked objects and MME perception on masked variants, against completion as mask dilation grows. The claim that over-redaction is cheap survives only if completion stays flat while utility stays flat.

Is it a leak or a hallucination? On the Visual Redactions Dataset, human-scored leakage of masked private attributes under black-box masking gives the real-world stake. If this leakage is near zero, the privacy half of the story fails even if hallucination rises.

Which rendering style works best? Across the five renderings, the labeled placeholder must reduce both completion bins relative to the black box, at no loss of visible-content utility, on at least two of three models.

Does the correction transfer? The COCO-mined steering vector must reduce leakage on VISPR and BIV-Priv-Seg without degrading visible-object POPE accuracy. The noisy in-domain vector tells us whether private data adds anything.

The single most decisive result is the comparison against prompt disclosure and generic hallucination steering on the leak bin. If redaction-specific steering does not beat both, the method is not needed, and the measurement becomes the paper.