AgentAug: Fake News Video Augmentation

• AgentAug is a data augmentation framework that simulates creative processes to generate diverse fake news videos for improved detection.
• It constructs multimodal material libraries and uses LLM-driven pipelines to recombine visual and textual elements into plausible fabricated narratives.
• An active-learning-based selection mechanism integrates only the most informative synthesized samples, thereby boosting detection performance.

AgentAug is a data augmentation framework for short video fake news detection that generates diverse fake news videos by simulating typical creative processes and then uses an active-learning-based selection mechanism to retain only synthesized samples that are useful for the current detector (Bu et al., 5 Oct 2025). It is motivated by a specific weakness in existing short video fake news detectors: they are primarily pattern-dependent, and the available training data capture only a sparse subset of the many-to-many relationships between source video material segments and fabricated news events. In the framework’s formulation, fake video detection should therefore not rely only on observed fake samples; augmentation should simulate the creation process of fabrication itself.

1. Problem setting and conceptual basis

AgentAug addresses short video fake news detection on platforms where a single video clip can be utilized in multiple ways to create different fake narratives, while a single fabricated event often combines multiple distinct video segments. Existing datasets such as FakeSV and FakeTT do not adequately reflect such relationships because collecting and annotating large-scale real-world fake news video data is difficult. The result is sparse coverage and non-comprehensive learning of the characteristics of potential fake news video creation (Bu et al., 5 Oct 2025).

The framework’s central insight is that existing detectors are not necessarily weak because of architecture alone; they are bottlenecked by sparse and insufficiently diverse training coverage. The paper argues that current pattern-based detectors mostly learn discriminative correlations from available training data, and that this biases them toward dataset-specific feature associations rather than generalizable principles of fake video fabrication. AgentAug therefore shifts augmentation from superficial perturbation toward simulation of plausible creator behaviors.

The motivating example in the paper uses the same source clip of simulated Ethiopian Airlines crash footage. Different fabrication routes can relabel the simulation as real black-box footage, insert unrelated clips to alter context, add a fictional allegation such as a deliberate crash over financial disputes, or repurpose the footage as coverage of the China Eastern Airlines crash. This example is used to illustrate the many-to-many logic that the framework is designed to cover.

2. Material library construction

The synthesis stage begins by constructing a material library from the training split only, explicitly to avoid test leakage (Bu et al., 5 Oct 2025). The textual material library is built from captions of real-world short news videos. The authors use Qwen-Max to analyze these captions and filter out obscure text with little event information. From the training portions of FakeSV and FakeTT they archive 2,106 Chinese and 910 English news text samples.

The visual material library is built by segmenting training videos into raw clips using TransNet-v2. These clips are then analyzed by Qwen-VL-Max, which produces objective content descriptions based on key visual elements and assigns each clip a type such as “real-shot video,” “real-shot photo,” “interview,” or “screenshot.” Qwen-VL-Max also rates each clip on visual quality, newsworthiness, and visual impact, each from 1 to 5 with justifications. After removing short or trivial clips, the final visual libraries contain 11,095 clips for FakeSV and 6,294 clips for FakeTT.

This construction step is not merely archival. It creates the substrate on which later fabrication workflows operate: text-only materials, visual-only materials, and paired multimodal materials. A plausible implication is that AgentAug treats augmentation as a library-driven recombination problem rather than as direct end-to-end video generation.

3. Fabrication taxonomy and LLM-driven pipelines

AgentAug explicitly focuses on textual and visual modalities, excluding audio and metadata. It describes fake creation as modality transformation over text only $(\mathrm{T})$, visual only $(\mathrm{V})$, or multimodal $(\mathrm{TV})$ materials. The framework formalizes four representative mappings:

$$\mathrm{V}\rightarrow \widetilde{\mathrm{T}\mathrm{V}}, \quad \mathrm{T}\rightarrow \mathrm{T}\widetilde{\mathrm{V}}, \quad \mathrm{TV}\rightarrow \widetilde{\mathrm{T}\mathrm{V}}, \quad \mathrm{TV}\rightarrow \mathrm{T}\widetilde{\mathrm{V}}$$

The first two are grouped as completion-based, and the latter two as manipulation-based (Bu et al., 5 Oct 2025).

Mapping	Fabrication category	Core operation
$$\mathrm{V}\rightarrow \widetilde{\mathrm{T}\mathrm{V}}$$	groundless fabrication	invent a fictional news-style textual narrative from visual material
$$\mathrm{T}\rightarrow \mathrm{T}\widetilde{\mathrm{V}}$$	misleading substitution	attach misleading visual evidence to a textual claim
$$\mathrm{TV}\rightarrow \widetilde{\mathrm{T}\mathrm{V}}$$	fact distortion	rewrite factual details while preserving visual plausibility
$$\mathrm{TV}\rightarrow \mathrm{T}\widetilde{\mathrm{V}}$$	selective editing	replace or insert video segments while preserving the textual frame

In the $$\mathrm{V}\rightarrow \widetilde{\mathrm{T}\mathrm{V}}$$ pipeline, one or more clips from the visual library are used as input. The LLM is prompted using the visual descriptions of those clips and generates a coherent news-style textual description or claim that fits the imagery while being fictional. The result is a fake news sample whose video remains grounded in real footage, but whose textual event is fabricated.

In the $$\mathrm{T}\rightarrow \mathrm{T}\widetilde{\mathrm{V}}$$ pipeline, the input is text only from the textual material library. AgentAug retrieves candidate visual clips by matching the text against clip descriptions using BGE embeddings. It then iteratively selects and appends relevant clips until a target duration is reached; the implementation sets the duration to 15 seconds. The augmented sample keeps the text claim but constructs a misleading video sequence from retrieved segments that appear relevant enough to support the text while not truly depicting the event.

In the $(\mathrm{V})$0 pipeline, the input is a real multimodal news sample. The LLM is prompted to modify factual details while preserving semantic consistency with the visuals. The output is a fake text-plus-original-video pair in which actors, causes, locations, or consequences may be altered while remaining visually believable.

In the $(\mathrm{V})$1 pipeline, the system first scores original video segments using visual quality, newsworthiness, and visual impact, then replaces low-scored segments with high-scored, non-overlapping segments from the library. Insertion or replacement is designed to introduce misleading visual evidence while preserving the textual frame.

Applied to the two benchmarks, these pipelines synthesize 3,278 augmented fake news videos for FakeSV and 2,592 for FakeTT.

4. Active sample selection and training integration

AgentAug does not train on all synthesized samples. Instead, it includes an active-learning-based sample selection strategy during training (Bu et al., 5 Oct 2025). After an initial detector is trained on the original human-annotated data, the procedure uses five-fold cross-validation on the training data. In each fold, four folds are used for training and one fold for validation.

At each iteration, the current model predicts probabilities on the validation set and computes uncertainty using entropy:

$(\mathrm{V})$2

where $(\mathrm{V})$3 is the predicted probability that a sample is fake. The top $(\mathrm{V})$4 fraction of validation samples with highest entropy are selected as anchors, with $(\mathrm{V})$5.

For each anchor, AgentAug retrieves the top-$(\mathrm{V})$6 most similar synthesized samples from the candidate augmentation pool using cosine similarity in the feature space extracted by the current detector. The paper states that $(\mathrm{V})$7 is chosen by validation search in $(\mathrm{V})$8 with step 100. These retrieved synthetic samples are temporarily added to the training set, and the detector is retrained for one iteration. Retention is decided by whether F1 on the anchor samples improves. If validation improves, the synthetic samples are retained; otherwise they are discarded or partly rejected.

This process continues for $(\mathrm{V})$9 rounds within each fold, where $(\mathrm{TV})$0 is chosen as the minimum number needed so all synthesized data can be considered once under the selected $(\mathrm{TV})$1. After all five folds finish, the framework aggregates how frequently each synthesized sample was selected across folds and keeps only those selected in at least three folds. The final model is then trained on the union of the original labeled data and this consistently selected synthetic subset.

The training enhancement is explicitly model-agnostic. The only requirement is that the downstream detector be trainable and produce prediction probabilities, since uncertainty is computed over predicted fake probability and retrieval uses the detector’s feature space. The paper evaluates this integration on FANVM, SVFEND, SVRPM, and MMVD, and states that the novelty lies in the data generation and adaptive sample selection procedure rather than in a modified detector architecture or task-specific loss.

5. Empirical results and ablation evidence

The evaluation uses two benchmark datasets with temporal splits of train/val/test = 70/15/15, following prior work, and the test set contains only human-produced videos (Bu et al., 5 Oct 2025). Dataset statistics are reported as follows: FakeSV has average duration 39.88s, 1,810 fake and 1,814 real samples, total 3,624; FakeTT has average duration 47.69s, 1,172 fake and 819 real samples, total 1,991. The metrics are accuracy, F1, precision, and recall.

On FakeSV, AgentAug with active learning improves MMVD from 75.83 Acc / 75.33 F1 to 77.12 / 76.69, FANVM from 78.41 / 77.89 to 81.37 / 80.42, SVFEND from 80.88 / 80.54 to 83.76 / 82.98, and SVRPM from 81.34 / 81.11 to 83.10 / 82.89. On FakeTT, it improves MMVD from 67.50 / 66.20 to 68.57 / 67.26, FANVM from 71.57 / 70.21 to 75.25 / 74.02, SVFEND from 77.14 / 75.63 to 80.43 / 78.61, and SVRPM from 81.79 / 79.42 to 82.86 / 80.57.

The paper reports that the strongest gains appear for weaker detectors, especially FANVM. This suggests that simpler models benefit more from richer and more diverse training signals, although stronger backbones also show stable gains.

The ablation studies further separate sample generation from sample selection. AgentAug$(\mathrm{TV})$2 is compared with AgentAug$(\mathrm{TV})$3, random sampling, and AgentAug$(\mathrm{TV})$4, class-balanced random sampling across the four fabrication types. These alternatives are described as unstable and sometimes harmful. For example, FANVM on FakeTT drops from 70.21 F1 to 66.77 with random sampling, and MMVD on FakeTT also declines under both random and balanced strategies relative to baseline. The active selection component is therefore presented as essential rather than optional.

A second ablation removes one fabrication category at a time using SVFEND. Full AgentAug reaches 83.76 Acc / 82.98 F1 on FakeSV and 80.43 / 78.61 on FakeTT. Removing any one category hurts performance: without $(\mathrm{TV})$5, FakeSV F1 is 81.10 and FakeTT F1 is 76.52; without $(\mathrm{TV})$6, FakeSV F1 is 79.24 and FakeTT F1 is 76.10; without $(\mathrm{TV})$7, FakeSV F1 is 81.93 and FakeTT F1 is 76.83; without $(\mathrm{TV})$8, FakeSV F1 is 81.82 and FakeTT F1 is 76.58. This is used to show that the four fabrication strategies are complementary.

The feature-space analysis uses t-SNE in the SVFEND feature space. Synthesized and human samples overlap substantially, which suggests that the augmented data are not out-of-distribution artifacts. At the same time, synthetic samples fill sparse regions in feature space, indicating that they expand the support of training data into plausible but previously underrepresented manipulation patterns.

6. Limitations, safeguards, and broader usage

The paper states several limitations explicitly (Bu et al., 5 Oct 2025). AgentAug does not synthesize fully generated fake videos from scratch; it only recombines and rewrites existing materials. The authors justify this by noting that fully generated fake news videos remain hard to control and are not yet a mature fabrication mode. They also note that the four-category taxonomy may not exhaust all possible perspectives on fake news video fabrication, and that the system uses task-specific agentic workflows rather than more autonomous self-improving agents that adapt generation based on feedback.

The limitations are accompanied by risks that are implicit in the design: LLM-generated narratives can hallucinate, synthetic examples can introduce noise, and generated fake-news-like content raises ethical concerns. The safeguard described in the paper is procedural rather than technical: synthetic samples are generated only for research, the authors state that they will not publicly release complete synthesized fake news videos, and reproducibility materials will be shared only under an application review mechanism.

The term “AgentAug” has also appeared in a broader, domain-shifted sense. In dermatology vision-language pretraining, “Derm1M-AgentAug” denotes an augmented training corpus produced by the MAGEN system, where low-quality image-text pairs are selectively repaired through foundation-model-assisted captioning, retrieval-based verification, and fallback handling for uncertain cases, yielding over 400k skin-image-text pairs (Li et al., 3 Dec 2025). This suggests that “AgentAug” has broadened from a single fake-news-video augmentation framework into a more general label for coordinated, agent-based data augmentation workflows that combine generation, retrieval, verification, and quality control.

In that broader sense, AgentAug refers less to a new backbone model than to a design principle: augmentation should simulate or repair the data-generating process in a task-specific, model-agnostic, and selectively filtered manner.