Three-language dev.to guides push DIY real-time deepfake detection ahead of elections

Three guides, one playbook

On October 3, 2026, dev.to's home feed carried three tutorials on building a deepfake detector, published within about two minutes of each other under the same author handle, leojulieta: one in Spanish, one in English, one in Portuguese. Each wraps the same pitch in a different toolchain — a browser extension, a Dockerised API and a hybrid scoring script — and each frames the project as a response to election-season synthetic media.

The coordination is part of the story. Community tutorials on deepfake detection have existed for years, but a same-day, three-language push aimed at journalists, creators and voters worried about fake candidate videos suggests demand for practical, local verification tooling has moved from research blogs into the mainstream developer feed.

What each guide builds

The Spanish-language guide targets the browser. It extracts frames with ffmpeg or the Canvas API, runs a fine-tuned XceptionNet model of roughly 30 MB, and applies temporal smoothing to suppress flickering false positives on short clips. A CLI script annotates each frame with a REAL or FAKE label plus a probability, and a companion Chrome/Edge extension runs the model in TensorFlow.js over any video element on a page, re-scoring every half second. Its practical advice: keep processing local for privacy, refresh model weights every three to four months, raise the decision threshold from 0.5 to 0.65 when low-quality clips trigger false alarms, and use an 8 MB TensorFlow Lite build on phones.

The English guide is written for newsrooms. It trains a MobileNet-V3 Small backbone — under 2 million parameters — with a temporal attention head, and reports 84% AUC on a held-out DFDC-2023 test set after five epochs on a single RTX 3060. The result ships as a Docker image exposing a Flask endpoint that returns a deepfake score and heat-maps; an ONNX export reaches about 25 fps on the same GPU. Its benchmark table is candid about limits: roughly 12 fps on an RTX 3060, 9 fps on an RTX 2070 and about 1 fps CPU-only — batch territory rather than live-stream filtering.

The Portuguese guide takes a hybrid route, blending a fast, CPU-friendly compression-artifact score computed from DCT coefficients with a lightweight 3D CNN, weighted 0.4 and 0.6 behind an adjustable 0.55 threshold, all inside a Docker container running TensorFlow 2.15. It also sketches a roadmap: audio-only voice-clone detection via Wav2Vec 2.0 spectrogram classification by late 2026, and INT8 quantisation for Raspberry Pi 4 deployment in 2027.

The motivating numbers, and where they clash

All three guides lean on alarming statistics, and they do not agree. The Spanish tutorial attributes a 73% rise in suspicious content on TikTok, Instagram and YouTube since January to Meta Trust & Safety, and cites a Berkman Klein Center figure of more than 1,200 fake videos detected 48 hours before votes in three democracies. The Portuguese tutorial instead quotes a DeepMedia Lab report claiming a 250% increase over twelve months, alongside dated 2026 incidents including a fake presidential-candidate clip that drew a reported 1.2 million retweets in three hours before Sensity AI flagged it. None of these figures is independently verifiable from the posts, and the two growth rates clearly measure different things. They work as motivation, not as evidence.

One further caveat for anyone following along: several links in the tutorials are placeholders. Both GitHub repositories use a yourname path, and the Portuguese guide's model download points to example.com, so readers will need to supply their own weights and datasets before anything actually runs.

Why it matters

The signal here is the packaging, not any single model. Detection research is being compressed into afternoon-scale projects — a browser overlay, a container endpoint, a hybrid script — published simultaneously in multiple languages, on the premise that free public APIs are too rate-limited or too scarce to trust during an election window. Self-hosted verification carries real costs: a detector at 84% AUC will still mislabel plenty of genuine footage, and the unverified statistics in the tutorials show how easily detection hype can echo the misinformation it targets. But if newsrooms and creators start adopting local scorers as first-pass triage, with human review behind them, these guides offer a plausible preview of how media verification gets distributed in 2026.