Effectiveness of EfficientNet-B4 and EfficientNet-B7 Deepfake Detection Techniques
PhD thesis — University of the Cumberlands, 2021. Full text: https://archive.org/details/jasmin-bharadiya-thesis
Evidence
Effectiveness of EfficientNet-B4 and EfficientNet-B7 Deepfake Detection Techniques
Deepfakes can fabricate a candidate saying things they never said; if detectors only reliably catch last year's fakes, the defense is permanently one generation behind the offense.
Deepfake detection is structurally reactive — detectors are trained on outdated generation tech; the thesis measures EfficientNet-B4/B7 in two phases (untrained on the latest deepfake datasets vs. trained on them, both benchmark-validated) to quantify how much latest-generation training data improves detection effectiveness.
Why it matters
Deepfakes can fabricate a candidate saying things they never said; if detectors only reliably catch last year's fakes, the defense is permanently one generation behind the offense.
Use when: A detector trained on yesterday's threat is a history lesson, not a shield — detection systems must be retrained on the newest generation of fakes or they measure the past.