Efficient Deepfake Detection for Low-Resource Settings

Problem

Deepfakes amplify misinformation, election manipulation, and fraud — and developing nations are especially exposed, with fragile information ecosystems and lower digital literacy. Yet the strongest detectors (large CNNs, Vision Transformers, optical-flow video models) assume high-end GPUs, big datasets, and constant connectivity. Those assumptions don't hold on the low-power, often offline devices common in these regions, leaving the most vulnerable populations the least protected.

Approach

This research surveys and compares detection methods through the lens of resource constraints, mapping each to its computational cost and suitability for edge deployment:

• H.264 motion-vector analysis — reuses motion data already inside the video codec to flag temporal inconsistencies cheaply, avoiding expensive optical flow.
• Steganalysis-based models — detect subtle pixel-level generation artifacts at a fraction of typical deep-learning cost.
• Multi-feature fusion — combines handcrafted features (HOG, LBP, KAZE) with classical ML classifiers for compact, smartphone-friendly models.
• Binary Neural Networks — 1-bit weights and activations for real-time detection on very low-power hardware.

It also weighs non-technical countermeasures — WhatsApp tiplines, community fact-checking, media literacy — that scale where pure-AI tooling cannot.

Results

• A comparative framework matching detection methods to accuracy/cost/resource-suitability trade-offs.
• Case-study evidence (India, Slovakia, Nigeria, Southeast Asia) showing that even simpler synthetic-audio manipulations are effective, so detection efforts must cover a broad spectrum, not just photorealistic video.
• A set of recommendations for governments, NGOs, technology developers, and communities centered on offline, locally owned solutions.

Learnings

• "State-of-the-art" accuracy is the wrong target when the deployment environment can't run the model.
• Generalization to unseen generation methods and robustness to compression/noise remain open problems.
• Sustainable impact comes from pairing efficient models with human-centered, low-cost verification channels.