TL;DR
CEVA (Complementary Evidence Verification Approach) is a novel framework for short-video fake news detection that reorganizes video frames into "claim-aligned" and "complementary-context" groups before they are fused with text. By explicitly preserving subtle contextual cues that might contradict a misleading claim—rather than just focusing on frames that match the text—CEVA achieves state-of-the-art performance across multiple datasets and languages.
What Problem Does This Paper Address?
Short videos are a potent medium for spreading misinformation because they can easily pair genuine, persuasive footage with completely unrelated or misleading claims. For example, a video of a staged emergency drill might be reposted as a "real-time crime in progress."
Existing automated detectors typically treat video frames as a uniform stream or use text-frame similarity to pick the "most relevant" frames. The authors argue this is a fundamental flaw: in a fake news video, the frames that most closely match the misleading text actually reinforce the lie. The "smoking gun" evidence that debunks the claim often resides in brief, low-similarity background shots—like a camera gimbal or a "behind-the-scenes" crew member—that traditional similarity-based filters would discard as noise.
Methods
The core of the CEVA framework is the Dual-Branch Complementary Reorganization (DBCR) module, which organizes evidence before multimodal fusion.
- Shot-Aware Sampling: Instead of uniform sampling, CEVA segments the video into shots and allocates a sampling budget to ensure dynamic segments and broad temporal coverage are preserved.
- Multi-View Semantic Profiling: Using a Vision-Language Model (VLM) embedding (Qwen3-VL), each frame is scored against five textual "views": title, claim, summary, details, and OCR text.
- Evidence Partitioning: Frames are ranked and split into two disjoint groups:
- $G_{align}$ (Explicit Alignment): Frames that strongly match the claim and title.
- $G_{comp}$ (Complementary Context): Frames that show high variance across different text views, likely containing the "diagnostic context" needed for verification.
- Two-Stage Textualization: The model generates global narrative descriptions ($T_1$) and specific, entity-aware descriptions for the reorganized frame groups ($T_2$).
- Fusion & Regularization: Visual features (from DINOv2) and text features (from Qwen2.5) are integrated via gated fusion and cross-attention. During training, a bidirectional contrastive loss (InfoNCE) is used to regularize the alignment between the video and text representations.
Datasets
The authors evaluated CEVA on two major short-video benchmarks:
- FakeSV: A Chinese dataset containing ~3,600 videos from platforms like Douyin and Kuaishou.
- FakeTT: An English dataset containing ~2,000 videos from TikTok. The use of an event-level split (ensuring no overlap of events between training and testing) prevents the model from simply memorizing specific news stories.
Key Insights and Ideas
- Similarity $\neq$ Veracity: High similarity between text and video is a poor proxy for truth in fake news detection. The "complementary" branch ($G_{comp}$) is essential because it captures the subtle contradictions that similarity-only methods ignore.
- Pre-fusion Organization Matters: Most models perform fusion on a "homogenous" stream of frames. CEVA proves that explicitly categorizing frames by their "evidential role" (supporting vs. contextualizing) before fusion leads to much cleaner feature representations.
- Transferability: The DBCR module is architecture-agnostic. The authors demonstrated that "plugging" DBCR into other existing detectors (like SV-FEND) significantly boosted their accuracy without changing their underlying fusion logic.
Key Results
CEVA set new performance benchmarks on both datasets:
- Accuracy: 87.98% on FakeSV and 86.67% on FakeTT, outperforming the previous state-of-the-art (ExMRD) by 1.08 and 2.39 percentage points, respectively.
- Ablation Success: Removing the $G_{comp}$ branch and its associated text descriptions caused accuracy to drop by over 3 percentage points, confirming the value of the "complementary" evidence.
- Zero-Shot Comparison: CEVA significantly outperformed large multimodal models (like GPT-4o-mini and Qwen2.5-VL) used in a zero-shot capacity, highlighting the need for task-specific training in misinformation detection.
Limitations and Future Work
The authors identify a specific failure mode: single-shot videos. If a video consists of a single continuous clip with no visual contradictions (e.g., a real explosion clip where only the location name in the caption is faked), CEVA’s reorganization strategy has no "contextual variation" to exploit. In these cases, the model may incorrectly predict "Real" because the footage perfectly matches the (misleading) title.
Future work may involve integrating external knowledge retrieval to verify entities (like specific locations or dates) that cannot be debunked through visual context alone.
Summary updated