Frequently Asked Questions
Questions and answers about detection accuracy, deployment, and legal admissibility for law enforcement and legal professionals.
How accurate is the detection engine?
Detection accuracy depends on synthesis method and quality. Our engine achieves >94% accuracy on state-of-the-art deepfakes (Deepfaceslab, First Order Motion Model), >88% on photosynthetic media (diffusion models, GAN-generated faces), and >96% on tampered video segments. Real-world performance varies by video codec, compression, and post-processing. We provide confidence scores on every detection, not binary verdicts, to support forensic analysis over enforcement.
What does the tool detect?
Synthetic media across video, audio, and still images. We detect face-swap deepfakes, generative AI faces (uncanniness artifacts), facial reenactment (emotion/expression transfer), lip-sync manipulation, voice cloning, and splice/composite editing in images. We also flag temporal inconsistencies (lighting, physics) frame-by-frame. We do NOT detect traditional watermarking or steganography.
What's the false positive rate?
~2-3% on authentic media, ~5-6% on heavily compressed video. Compression artifacts, low resolution, and unusual lighting can trigger false alerts. We mitigate this by reporting confidence scores (0-100) rather than binary verdicts, and by providing visual heatmaps of detected anomalies. We recommend human review of alerts above 65 confidence for legal casework.
For law enforcement: False positives exist. Our tool narrows investigative focus; it does not replace forensic expertise or chain-of-custody procedures.
Is the output admissible in court?
Our report is a forensic aid, not expert testimony. Admissibility depends on jurisdiction, case law, and expert qualification. We provide detailed analysis logs (frame-level detections, confidence scores, model version) to support expert witness testimony. We do NOT provide legal opinions. Defense counsel can challenge methodology; we publish our architecture and training data sources to enable independent validation.
How does deployment work?
On-premises, air-gapped, or via secure API. LEA and legal teams can deploy the inference engine on internal servers (no cloud upload). All processing stays local. For smaller departments, we offer a secure API with TLS 1.3 and end-to-end encryption. Video evidence never leaves your infrastructure without explicit approval.
What file formats are supported?
Video: MP4, MOV, AVI, MKV, WebM (H.264/H.265 codecs). Audio: MP3, WAV, AAC, FLAC. Images: PNG, JPEG, TIFF, WebP. Max file size: 2GB video, 500MB audio, 100MB per image. Processing time: ~1-2x realtime for video (1-hour video = 1-2 hours analysis).
Can I integrate this with our existing case management system?
Yes, via REST API or batch processing. We provide JSON output with frame-level detections, timestamps, and confidence scores. Integration with CaseManager, Axon, and LEAP is available; custom integrations can be built for specialized systems. Bulk processing (batches of 100+ videos) is supported.
What training data is used?
Synthetic media from public benchmarks (FaceForensics++, DFDC, WildDeepfake) plus in-house law enforcement casework with consent. We do NOT train on real victim/evidence data. Model cards are published for each version, including precision/recall curves by synthesis method. Training is frozen between releases; no continual learning on your case files.
How often is the detection engine updated?
Quarterly. As adversaries develop new synthesis techniques, we release new model weights. You control when to upgrade; we maintain backwards compatibility for analysis logs so historical cases remain interpretable. Major methodology changes are announced 60 days in advance.
What is the per-case cost?
Tiered by department size and annual case volume. Small departments (<50 cases/year): $199 per case. Medium (50-500): $99 per case. Large (500+): contact sales for volume licensing. On-premises deployment: $15k/year per seat. No per-case fees for on-premises.
Is there a trial or evaluation period?
Yes. 30-day evaluation license with up to 10 case analyses. No credit card required for LEA and legal firms. Evaluation includes full API access, documentation, and 1:1 onboarding call.