[ 01 ]
Why looking and listening is not enough
Modern face-swap, lip-sync and voice-clone tools produce output that passes a casual review. Reviewers disagree with each other, and they cannot explain a judgement made by eye in a way that survives a challenge. Detection has to produce measurements that someone else can repeat.
[ 02 ]
Video: temporal and encoding evidence
Real footage changes smoothly from frame to frame. Manipulated regions, especially around faces, can flicker in brightness, edge sharpness or motion. Encoders also leave a pattern in how frames are grouped and quantised, and an edited segment can carry a different pattern from its neighbours.
Platform re-uploads rewrite compression, so a good report states how much confidence the file's history allows.
[ 03 ]
Images: recompression, noise and structure
Error level analysis, noise residual checks and copy-move searches highlight regions that behave differently from the rest of the image. A camera sensor leaves a faint noise signature, and pasted or generated regions rarely match it. Embedded metadata and thumbnails are cross-checked against what the pixels show.
[ 04 ]
Audio: continuity, noise floor and encoding
Spliced and synthetic speech can interrupt spectral continuity, change the background noise floor or show re-encoding traces. Where a mains-hum signal is present, its variation over time can reveal joins. Phone codecs remove detail, so the channel assumptions belong in the report.
[ 05 ]
Combine sensors and keep a record
Each sensor has blind spots. Combining independent sensors, showing where they disagree and recording the parameters used gives an examiner a result that can be repeated and explained. This is the approach SpearTrace takes.
Quick answers
Can deepfakes always be detected?
No method is perfect. Detection depends on file quality, compression and how the media was made. Forensic tools report what they measured and what they cannot determine, so humans decide with evidence.
What is the difference between AI detectors and forensic sensors?
AI detectors return a learned probability that can be hard to explain. Classical forensic sensors measure specific physical or structural properties, so each finding points to evidence an examiner can inspect.