Unit 9. Deepfakes and synthetic media

Synthetic images, video and voice have become tools of fraud and deception, and security teams are asked both to detect them and to stop the frauds they enable. Lesson 9.1 explains how synthetic media is made, from the shared-encoder autoencoders behind the original face-swap deepfakes to generative adversarial networks, diffusion models, real-time face reenactment and voice cloning, and what each method demands of an attacker. Lesson 9.2 covers detection by spectral artifacts and by learned classifiers, and the evidence that detectors often fail on generators and datasets they were not trained on, lose accuracy and shift their scores under compression, can be evaded, and produce mostly false alarms at realistic base rates. Lesson 9.3 turns to approaches that work from the source, content provenance with C2PA Content Credentials and watermarking of generated text and images, with their limits, and then to the documented frauds with cloned voices and fabricated video calls and the verification procedures that stop them.

After this unit you will be able to explain the generation pipelines and the traces they leave, compute and interpret azimuthally averaged spectra, measure a detector's generalization across generators and its behavior under compression, compute the precision of a detector at a realistic base rate, state precisely what a signed provenance manifest and a watermark do and do not establish, and design payment and approval procedures that do not depend on recognizing synthetic media. The unit builds on Lesson 1.2 (base rates and precision), Unit 6 (convolutional networks) and Unit 7 (evasion attacks). Allow about 8 to 10 hours for the three lessons, Lab 9, the exercises and the quiz.

Lessons

1. 9.1 How synthetic media is made
2. 9.2 Detecting synthetic media
3. 9.3 Provenance, watermarking and fraud response

Labs

Practice

• Exercises with worked solutions
• Self-check quiz