A misleading social-media post can be faked in many ways. Someone can edit the photo, rewrite the caption, swap a named entity, flip the sentiment, or pair a genuine photo with a false story. Current benchmarks now catalog 11 to 12 such manipulation types, yet the detectors trained on them either predict a single "fake" label or share one network across three or four coarse sub-tasks. When a detector is fine-tuned on new manipulation types, it forgets the old ones. On posts that combine several manipulations at once, everything breaks down: humans identify the source of misinformation about 38% of the time and zero-shot GPT-4V reaches about 51% F1 on MMFakeBench. The multi-task learning literature has a ready explanation for this pattern. Objectives that pull shared weights in opposite directions interfere, and the interference grows with the number of tasks. Nobody has checked whether manipulation types actually behave this way.
We propose to measure, layer by layer, where and in which direction the training signals for different manipulation types conflict inside a vision-language detector. Our hypothesis is that the conflict is not spread evenly. It concentrates in the late layers of the image encoder, between tasks that need low-level pixel evidence (edited images) and tasks that need meaning-level evidence where pixel detail is noise (image–text mismatch). Wherever the measurements show conflict, we then give the detector separate low-rank adapters, which are small trainable weight add-ons. A learned gate chooses which adapters to use from the input alone, so no manipulation label is needed at test time. In two sentences: find where manipulation types damage each other inside the network, and split capacity only there. If the conflict is not there, the simple pooled detector is enough, and we will say so.
Core Claim
The natural unit of "task" in multimodal misinformation detection is the evidence channel that carries the manipulation, not the dataset's label. Concretely, the channels are pixel-level forgery, textual distortion, cross-modal inconsistency, and contradiction with external knowledge. Training signals from different channels conflict in specific layers. The conflict is asymmetric: one task can help another without being helped back. And it is the mechanism behind two documented failures, forgetting when new fake types are added and collapse on mixed-source fakes.
The asymmetric, layer-local conflict structure can be measured. Allocating adapters according to that measurement should beat three alternatives at equal parameter budget: pooling all types into one adapter, giving every type its own adapter, and the spectral regularization of Scalable Multi-Task Low-Rank Model Adaptation (mtLoRA).
The physical intuition behind the hypothesis is simple. Encoders pretrained to align images with captions are designed to ignore small pixel changes. That indifference helps judge whether a caption fits a photo, but it is exactly wrong for spotting a spliced face.
Relation to Prior Work
Three misinformation papers already use multi-task structure, and they motivate this proposal by showing the gap.
MMFND uses a progressive layered extractor with hand-assigned text, image, and cross-modal expert groups. It down-weights the cross-modal branch when the similarity between image and text features is low. Its three sub-tasks and its expert allocation are fixed in advance.
MMDFND treats nine topical domains as tasks with domain-specific experts. It needs the domain ID at test time, and it models only real-versus-fake.
SDML trains four outputs on its MFND dataset: real/fake, image forgery, text forgery, and bounding-box localization. It does this without quantifying interference. It also does not use the 11 manipulation types annotated in its own data as separate objectives.
Our proposal differs from all three in two ways. The task partition is decided by measured conflict rather than by design, and the decision is made per layer.
Three multi-task scaling methods supply our technique, and each constrains it.
AdapterShare starts from one shared adapter and splits it when the cosine similarity of validation gradients turns negative. We reuse that splitting rule. We replace its symmetric conflict measure with a directional one, addressing the limitation the authors themselves name.
mtLoRA reports two things relevant here. Standard multi-task LoRA collapses from 88.2% to 2.0% on DOTA when going from 5 to 15 tasks. The softmax in attention also acts as a conflict amplifier, with a 60.6% conflict rate that drops to 53.6% when adaptation moves to the feed-forward blocks. We do not assume the collapse transfers to manipulation types; we test it. We use mtLoRA as the strongest scaling baseline and test its attention-versus-feed-forward finding directly.
HarmoDT learns task-specific masks over parameter subsets where tasks agree. It informs our gate, but it was built for reinforcement-learning policies that know their task ID. Our gate cannot rely on one.
On the misinformation side, MMFakeBench both motivates and challenges this work. It reports that fine-tuning a specialized detector on new manipulation types causes severe forgetting, and it concludes that prompted vision-language models with search tools are the better route. Our proposal tests that conclusion directly rather than assuming it.
The Knowledge-guided Dual-inconsistency Network separates cross-modal inconsistency from content-knowledge inconsistency. It gives us the knowledge-contradiction channel, but trains both channels with a single shared optimizer.
FineFake offers a six-way taxonomy that maps cleanly onto our channels. Its adversarial training strips out domain information. We consider that a design choice our measurements can evaluate, not one we adopt.
Research Program
Stage 1: conflict mapping. The first deliverable is a conflict map. We take a vision-language detector and attach low-rank adapters to every attention and feed-forward block of both encoders and the fusion layers. The backbone is a CLIP-style image and text encoder joined by cross-attention fusion, matching the detectors evaluated in SDML and MMFakeBench. Each manipulation type gets its own lightweight output head: the 11 MFND types, the DGM⁴ categories, and the five FineFake fake types.
During training we record, for every layer $l$ and every ordered pair of types $(i, j)$, a directional affinity:
$$ A^{(l)}_{i\to j} = 1 - \frac{L_j\big(\theta - \eta\, g^{(l)}_i\big)}{L_j(\theta)} $$
The symbols mean the following:
- $\theta$ is the current adapter weights.
- $g^{(l)}_i$ is the gradient of type $i$'s loss restricted to layer $l$'s adapter.
- $\eta$ is a small lookahead step.
- $L_j$ is type $j$'s loss on a held-out batch.
A negative value means that a step taken for type $i$ hurts type $j$. Because $A_{i\to j}$ need not equal $A_{j\to i}$, the measure captures asymmetry that a gradient cosine cannot.
The map tests three predictions:
- Pixel-forgery types conflict with cross-modal-mismatch types mainly in the late image-encoder layers.
- Textual-distortion types share comfortably with one another.
- Conflict is higher in attention adapters than in feed-forward adapters, as mtLoRA found for vision tasks.
Stage 2: conflict-guided allocation. We start from one shared adapter per layer. When the averaged negative affinity between groups of types crosses a threshold chosen on validation data, we split that layer's adapter. Following AdapterShare, the split requires the conflict to persist across several measurement windows, so that noise does not trigger it.
Asymmetry changes what a split looks like. If knowledge-contradiction training helps cross-modal consistency but not the reverse, the split lets the cross-modal branch keep reading the shared adapter while blocking its gradients from flowing into it.
At test time, a small gate reads pooled features from the shared layers and outputs soft weights over the branches at each split layer. The gate is trained with the type labels as auxiliary supervision, then evaluated with no labels. This matters because mixed-source fakes need several branches at once.
Stage 3: continual protocol. We add manipulation families in sequence, for example textual distortions first, then image edits, then cross-modal swaps. When a family arrives, its conflict with existing branches decides whether it gets a new branch or joins an existing one. Existing branches whose affinity with the new family is negative are frozen. This directly addresses the forgetting MMFakeBench reports.
Baselines
Naive reference points. The first reference is the do-nothing point: the frozen backbone with a linear real/fake probe. Above it sits the pooled detector: one shared LoRA adapter at every layer, trained on all manipulation types with a single multiclass head and class-balanced sampling. The pooled detector is the baseline most likely to embarrass the proposal. If manipulation types do not really conflict, nothing should beat it, and our Stage 1 map would show why.
Strongest trained method on the training benchmarks. This is SDML from the MFND paper, re-run with the same backbone. The file reports that SDML outperforms its baselines on AUC, accuracy, mean F1, and localization IoU on MFND and DGM⁴. It exceeds baselines by 3.51% IoU-mean and 3.14% IoU50 in image localization. No absolute SDML accuracy or F1 on MFND was found on Lacuna, so we will reproduce it ourselves and report our reproduced number alongside the published relative gains.
Strongest scaling method. This is mtLoRA, with its spectral regularization, feed-forward-block adaptation, and per-dimension routing. It is the published method most directly aimed at many-task adapter collapse.
Strongest method on the held-out mixed-source test. This is GPT-4V with the MMD-Agent tool-use procedure, which reaches 61.6% F1 on MMFakeBench according to MMFakeBench. We will also run LLaVA-1.6-34B with MMD-Agent as an open model; the same paper reports it at 49.9% F1.
Specialist baseline. We train one adapter per manipulation type. Selection is done either by an oracle that knows the type or by the same learned gate we use.
Fairness. All trained methods use the same backbone, the same training data, and the same total adapter parameter count. Ranks are adjusted so that pooled, specialist, mtLoRA, and our model match. Every method gets the same number of hyperparameter trials, and thresholds are tuned only on MFND and MMFakeBench validation splits.
Ablations. Two ablations isolate the key ingredient. In the first, splits are placed by type label at every layer rather than by measured conflict. This tests whether measurement matters or splitting alone helps. In the second, splits use symmetric gradient cosine rather than directional affinity. This tests whether asymmetry matters.
What counts as winning. Beating the pooled detector, SDML, and mtLoRA by a margin larger than the seed-to-seed spread on MFND mean F1 and localization IoU. On MMFakeBench, exceeding the 61.6% F1 of the agent pipeline with a single forward pass of a model far smaller than GPT-4V.
Why not just…?
Why not just pool every manipulation type into one fine-tuned adapter? This is a real threat. The 88.2%-to-2.0% collapse in mtLoRA came from aerial detection tasks with different outputs. Manipulation types share one input distribution and may simply be classes of a single problem. The Stage 1 map settles it. If cross-channel affinities are not negative in any layer, our method should place no splits and reduce to the pooled detector. We would then report that the multi-task framing of misinformation is unnecessary at this scale, which is itself a useful negative result. The proposal only claims an advantage on the types and layers where measured conflict exists.
Why not just give each manipulation type its own adapter? Per-type adapters avoid interference, but they also block sharing among related types. MMDFND reports that removing cross-domain sharing hurts data-poor domains. Per-type adapters also need either a known type or a gate. Mixed-source fakes are by definition several types at once, and a gate choosing among 11 isolated adapters has to recombine evidence it was never trained to combine. The oracle-specialist baseline bounds what isolation can achieve. If our model matches it with fewer parameters and beats the gated version on MMFakeBench, sharing earned its place.
Why not just use experience replay to stop forgetting? Replaying a buffer of old examples during sequential training is a strong, simple fix. If replay alone removes the forgetting MMFakeBench describes, our continual-protocol claim collapses to parameter efficiency. We therefore include pooled-adapter-plus-replay at matched buffer size in Stage 3. We expect replay to reduce forgetting but not eliminate it where conflict exists, because replayed old-type gradients still fight new-type gradients in the same weights. If replay wins, we will say so.
Why not just prompt a large vision-language model with search tools? MMFakeBench argues for exactly this. Its best pipeline, GPT-4V with MMD-Agent, is also the strongest number we know of on mixed-source fakes, at 61.6% F1. The same paper notes that search fails on breaking or niche news. Pixel-level edits do not show up in search at all. Running a multi-step tool loop on every post is also expensive. Our MMFakeBench comparison measures the trade directly, reporting F1 per distortion category alongside cost per post. We expect a trained detector to win on visual and cross-modal distortions and possibly lose on textual-veracity distortions that need world knowledge.
Proposed Experiments
Training and evaluation data. We train on MFND's 95k training samples with 11 manipulation types and on DGM⁴. We evaluate on the MFND and DGM⁴ test splits using AUC, accuracy, mean F1, IoU-mean, and IoU50. MMFakeBench serves as a held-out mixed-source test. Its 1,000-sample validation split is used only for threshold and gate calibration, and we report F1 overall and per distortion category on its 10,000 test pairs. FineFake adds the knowledge-contradiction channel and topic and platform shift; there the comparison point is KEAN's binary F1 of 0.797 reported in FineFake.
Scaling curve. We train every method on 3, 6, 11, and finally all available manipulation types across the combined data. If the pooled detector's per-type F1 falls with task count while ours stays flat, interference is real and our allocation addresses it.
Decisive checks. Three outcomes would decide the proposal:
- The conflict map must show negative directional affinity between pixel-forgery and cross-modal types concentrated in late image-encoder layers. It must also show higher conflict in attention than in feed-forward adapters. If neither appears, the mechanism is wrong.
- The measured-conflict model must beat the label-split and symmetric-split ablations at equal parameters. Otherwise the measurement contributes nothing beyond extra capacity.
- In the sequential protocol, forgetting on earlier families, measured as the drop in their F1 after later families are added, must be smaller than with pooled training plus replay. Its gain on MMFakeBench mixed-source fakes must also exceed the gain on single-source fakes, since mixed fakes are where using several branches at once should matter most.
References
- MMFakeBench: A Mixed-Source Multimodal Misinformation Detection Benchmark for LVLMs
- Multi-Task Network Guided Multimodal Fusion for Fake News Detection
- MMDFND: Multi-modal Multi-Domain Fake News Detection
- Multimodal Fake News Detection: MFND Dataset and Shallow-Deep Multitask Learning
- AdapterShare: Task Correlation Modeling with Adapter Differentiation
- Scalable Multi-Task Low-Rank Model Adaptation
- HarmoDT: Harmony Multi-Task Decision Transformer for Offline Reinforcement Learning
- FineFake: A Knowledge-Enriched Dataset for Fine-Grained Multi-Domain Fake News Detection