The most common kind of visual misinformation does not involve doctored pixels. It is a real photo with a false caption: a genuine picture of a politician at a rally, captioned as their arrest. Detectors that compare image and caption tend to fail on exactly these cases. They see the politician's face, read the politician's name, register a strong match, and declare the pair genuine, without checking whether the setting, objects, or event in the photo agree with the rest of the caption. This proposal treats that failure as shortcut learning, which a separate literature already handles in vision: during training, find the evidence the network relies on most, hide it, and require a correct answer anyway.
The new ingredient is how the hiding is done. In image–caption consistency checking, hiding evidence in only one modality changes the correct label. If you black out the face but leave the name in the caption, a genuine pair now looks mismatched. We propose paired erasing: find the caption phrase and image region that agree most strongly, remove both together, and train the detector to reach the same verdict from what is left. Removing an agreeing pair from both sides leaves a genuine pair genuine and a repurposed pair repurposed. That property is what lets adversarial erasing carry over to misinformation detection without adding label noise.
Core Claim
Our claim is that out-of-context detectors fixate on the single strongest image–caption correspondence, usually a person or a named entity. Training with jointly erased, label-preserving views will reduce this fixation more than random erasing, more than one-sided attention erasing, and more than simply masking every detected face and name. We expect the largest gains where the shortcut is most tempting: the Person Match split of NewsCLIPpings, and test sets whose people never appear in training.
Where This Sits
The masking side supplies the method. Self-challenging Improves Cross-Domain Generalization zeroes the feature dimensions with the largest loss gradients and trains on what remains. It reports that this beats random dropout and lets a ResNet-50 roughly match a ResNet-101 on ImageNet. Its single-modality, label-agnostic masking is what we must change for a pairwise consistency task. Revisiting Structured Dropout and Top-DB-Net give the same lesson in other settings: masking the high-activation regions specifically spreads what the network attends to.
The closest prior method is Improving Referring Expression Grounding With Cross-Modal Attention-Guided Erasing. It erases either the most-attended word or the most-attended image region to force a grounding model to find other correspondences. It raised RefCOCOg test accuracy from 78.12% to 80.37% over its MattNet baseline. That paper works only on pairs that are known to match, and erases one side at a time. It also warns that erasing a key word in a short phrase can destroy the phrase's meaning. Both facts motivate our design: on a consistency task, one-sided erasing is exactly the operation that breaks labels.
The misinformation side supplies the evidence for the gap and the evaluation. NewsCLIPpings builds repurposed pairs under several matching strategies. It reports CLIP and VisualBERT at 60–70% accuracy and humans at 65.6% overall and 35% on falsified samples, with Person Match among the hardest settings. Generalizing to the Future shows that text-only fake news detectors memorize entity–label correlations that collapse over time. It tries random entity masking, and leaves visual entities explicitly open. We extend its diagnosis across modalities, and we replace uniform masking with masking guided by the model's own strongest correspondence.
Interpretable Multimodal Out-of-Context Detection with Soft Logic Regularization breaks captions into phrases and enforces the rule that a pair is genuine only if every phrase agrees with the image. It reports about a three-point gain over CLIP and VisualBERT. Its rule is compatible with ours, and our erased views are a way to make the model actually check the phrases other than the dominant one. LOGRAN never perturbs the image side.
Retrieval-based systems change the question rather than answering it. Examples are Support or Refute, which clusters retrieved evidence by stance, and MAD-Sherlock, which runs multi-agent debate over web search. They bring in outside facts, and they fail when search returns nothing useful, as MAD-Sherlock acknowledges. Two findings constrain our design. First, Fact or Fake? shows that feeding pixel-forensics verdicts into fact-checkers lowers F1 by 0.04–0.08, because a "REAL" label anchors the model. So we do not rely on forensic signals. Second, Inconsistency Matters separates shared and modality-specific features to expose mismatch. That supports the intuition that the shared, matching evidence is what fools the model.
Research Program
The detector is a frozen CLIP image and text encoder topped by a small trainable cross-attention head. The head takes word tokens and image patch tokens and outputs a genuine-versus-repurposed score. From the head we compute an alignment matrix $A_{ij}$: the attention weight between caption token $i$ and image patch $j$, multiplied by the gradient of the loss with respect to that weight. We use this product because the self-challenging work found gradients a better guide to what the model actually uses than activations alone.
We group tokens into phrases with a noun-phrase chunker and group patches into connected regions. Then we pick the set $S$ of phrase–region pairs with the highest summed alignment, up to a fraction $p$ of total alignment mass. Both members of each selected pair are removed: the phrase is replaced by a mask token and the region's patch tokens are dropped. Writing $x \setminus S$ for this jointly erased input, the training loss is
$$ \mathcal{L} = \mathrm{CE}(f(x), y) + \lambda\, \mathrm{CE}(f(x \setminus S), y), $$
where $f$ is the detector, $y$ the genuine/repurposed label, CE is cross-entropy, and $\lambda$ weights the erased view.
Over-erasing is a real risk, for example a caption like "Obama speaks" that consists of nothing but the matched pair. We handle it with an abstention rule. If the alignment mass left after erasure is below a threshold, the erased view gets no loss for that example. This is a semantic constraint of the kind the research file's third gap calls for, and it avoids forcing the model to rule on empty evidence.
The analysis goes beyond accuracy. For each trained model we measure two reliance quantities, both on held-out genuine pairs. The first is how much of the head's attention lands on face and named-entity regions. The second is how often the verdict flips when only the face is occluded. A successful method should move attention to context while keeping its verdicts stable under face occlusion.
Baselines
The naive reference is the same CLIP-plus-head detector trained with plain cross-entropy. All variants share its encoder, data, number of training steps, and a tuning budget of the same number of runs over $\lambda$ and the masking rate.
The first masking baseline is random erasing of image patches and caption tokens at a matched rate. It tests whether targeting matters at all.
The second, and the most important, is fixed entity masking. A face detector and a named-entity tagger remove faces and person names jointly, both sides at once. It is a paired, heuristic version of the entity masking in Generalizing to the Future. It needs no gradients and is the cheapest serious competitor.
The third is a one-sided port of CM-Att-Erase with identical selection but erasure on one modality only. This is the ablation that isolates our key ingredient, joint label-preserving removal. A fourth is feature-level muting in the style of RSC on the fused representation.
Among published fine-tuned, retrieval-free systems, LOGRAN reports roughly three points over its CLIP and VisualBERT baselines on NewsCLIPpings. We will reimplement it on our backbone and also combine it with paired erasing.
For context only, retrieval systems report higher numbers. SEN reaches 87.1% NewsCLIPpings accuracy (Support or Refute). MAD-Sherlock with GPT-4o reaches 90.8% on NewsCLIPpings and 85.2% on VERITE, against Sniffer at 88.4% and CRAVE at 85.0% on NewsCLIPpings as reported in MAD-Sherlock. These systems use external evidence, so we compare against them only in a separate retrieval-augmented track, described below.
Beating the baselines means all of the following. Overall accuracy rises over the plain and LOGRAN detectors. Person Match falsified accuracy rises by more than the run-to-run spread across five seeds. Both one-sided erasing and fixed entity masking are beaten on the entity-held-out split and on VERITE.
Why not just…?
Why not just use random erasing or dropout? The self-challenging and structured-dropout papers report that blind masking leaves enough of a dominant feature intact for the network to keep using it. That evidence comes from image classification, not misinformation. The random-erasing baseline at a matched rate settles the question here. If it equals paired erasing, targeting adds nothing in this setting.
Why not just mask every face and name? This is the real threat. If the entity shortcut is the whole story, a face detector plus a tagger may do as well at lower cost. We expect the two methods to differ where the dominant match is not a person. The Scene Match splits of NewsCLIPpings, where Support or Refute reports its largest scenario gain, and the cross-modal consistency subset of MMFakeBench contain objects, logos, and landmarks that a face detector never removes. If fixed masking matches paired erasing on those splits too, we will report that the cheaper heuristic suffices and that model-guided selection is unnecessary.
Why not just use a larger vision-language model or a retrieval agent? MMFakeBench reports GPT-4V at about 51% F1 with standard prompting and LLaVA-1.6-34B at 25.7%, so scale alone does not close the gap. Retrieval agents score far higher but depend on what search returns. Our claim is about the internal comparison that every pipeline, including retrieval ones, eventually performs. We test this directly with a retrieval-augmented track. In that track, the SEN-style evidence comparison step is run once with the plain detector and once with the paired-erasing detector, so we can see whether the gain survives once outside evidence is available.
Proposed Experiments
Training uses NewsCLIPpings (NewsCLIPpings). We report accuracy per matching scenario, with genuine and falsified accuracy reported separately.
The first decisive check is an entity-held-out split. We partition NewsCLIPpings by depicted person so that no test identity appears in training. If paired erasing removes reliance on familiar faces, its advantage should grow on this split relative to the standard split. Fixed entity masking should help here too, so this split mainly separates both of them from the plain model.
The second check is transfer without retraining. The NewsCLIPpings-trained models are evaluated on VERITE, which is built so that neither modality alone suffices, and on the cross-modal consistency distortion portion of MMFakeBench.
The third check is mechanistic, using the two reliance quantities from the research program. Paired erasing should lower attention on entity regions and make verdicts more stable under face occlusion. If accuracy rises but the reliance measures do not move, the shortcut account is wrong and the gain has some other cause. The one-sided ablation should show label damage directly: one-sided erasing of genuine pairs should create apparent mismatches, visible as a higher rate of genuine pairs flipped to repurposed.
We also sweep the erasure fraction $p$ and the abstention threshold, and report the extra cost of the backward pass needed for gradient-weighted alignment. Together these runs tell us whether the joint, label-preserving erasure is what matters, or whether a cheaper heuristic captures the same benefit.
References
- Self-challenging Improves Cross-Domain Generalization
- Revisiting Structured Dropout
- Top-DB-Net: Top DropBlock for Activation Enhancement in Person Re-Identification
- Improving Referring Expression Grounding With Cross-Modal Attention-Guided Erasing
- NewsCLIPpings: Automatic Generation of Out-of-Context Multimodal Media
- Generalizing to the Future: Mitigating Entity Bias in Fake News Detection
- Interpretable Multimodal Out-of-context Detection with Soft Logic Regularization
- Support or Refute: Analyzing the Stance of Evidence to Detect Out-of-Context Mis- and Disinformation
- MAD-Sherlock: Multi-Agent Debate for Visual Misinformation Detection
- Fact or Fake? Assessing the Role of Deepfake Detectors in Multimodal Misinformation Detection
- Inconsistency Matters: A Knowledge-guided Dual-inconsistency Network for Multi-modal Rumor Detection
- MMFakeBench: A Mixed-Source Multimodal Misinformation Detection Benchmark for LVLMs