Spaces:
Sleeping
Sleeping
| title: DEEPFYND Intelligence Plane | |
| emoji: π | |
| colorFrom: green | |
| colorTo: gray | |
| sdk: gradio | |
| app_file: app.py | |
| pinned: false | |
| license: mit | |
| short_description: Explainable deepfake detection API for the DEEPFYND platform | |
| sdk_version: 6.20.0 | |
| # DEEPFYND β Intelligence Plane | |
| The analytical core of the DEEPFYND platform (MSc Digital Forensics & Cybersecurity). | |
| Serves both a demo UI and the REST API consumed by the platform's web, Android and | |
| Telegram channels. | |
| ## API contract | |
| `POST /analyze` | |
| ```json | |
| { "file_url": "https://...", "media_type": "image|audio|video", "scan_id": "rec..." } | |
| ``` | |
| Response: | |
| ```json | |
| { | |
| "verdict": "authentic|suspicious|likely_deepfake", | |
| "confidence": 87.4, | |
| "insights": "β’ ...", | |
| "heatmap_url": "https://.../heatmaps/rec....jpg", | |
| "file_hash": "sha256:...", | |
| "raw_scores": {} | |
| } | |
| ``` | |
| `GET /health` β liveness check. | |
| `GET /heatmaps/{name}` β serves generated explanation heatmaps. | |
| ## Layered architecture | |
| - **Layer 1 β Media forensics:** EXIF/metadata inspection, generator-typical dimension checks. | |
| - **Layer 2 β AI detection:** ViT image detector; wav2vec2 audio detector; frame-sampled video. | |
| - **Layer 3 β Risk scoring:** three-band verdict (authentic / suspicious / likely deepfake) with confidence. | |
| - **Layer 4 β Explainability:** gradient-based saliency heatmaps and plain-language reasons. | |
| ## Fail-honestly design | |
| If media cannot be fetched, decoded, or analysed, the service returns an HTTP error. | |
| It never fabricates a verdict. The upstream orchestration then records `Status=failed` | |
| with a plain-language message, and users are told nothing was assessed. | |
| ## Configuration | |
| Set in **Settings β Variables and secrets**: | |
| - `SPACE_URL` = the public URL of this Space | |
| (e.g. `https://lericsg-deepfynd-intelligence.hf.space`). | |
| Required so that generated heatmap URLs are publicly addressable. | |
| ## Baseline models | |
| This deployment uses publicly available pretrained detectors as a **baseline**: | |
| - Image/video: `prithivMLmods/Deep-Fake-Detector-v2-Model` | |
| - Audio: `Heem2/wav2vec2-based-Deepfake-Audio-Detection-V3` | |
| Fine-tuning on Ghana-relevant data is scoped as future work. | |
| ## Disclaimer | |
| Results are decision support, not proof. Verify important content with a professional fact-checker. |