Artificial intelligence forensics

• Uvod: AI forenzika kao disciplina
• Kontekst, motivacija, EU AI Act
• Arhitektura modernih AI sustava: LLM-ovi, difuzijski modeli, GANs
• Digitalni trag AI sustava: metapodaci, log datoteke, API pozivi
• Detekcija AI generiranog teksta: perplexity, burstiness, watermarking
• Forenzika promptova i konverzacijskih sustava
• Prompt injection napadi
• Deepfake – tehničke osnove: GAN-ovi, face-swap, voice cloning
• Detekcija deepfake videa i slika: CNN klasifikatori
• FaceForensics++ Forenzika audio sadržaja: kloniranje glasa, spektralna analiza
• Forenzika AI generiranih slika: C2PA standard, content provenance
• Forenzika ML modela: model stealing, membership inference, watermarking
• Forenzika podataka za treniranje: data poisoning, backdoor napadi
• Objašnjivost AI odluka: LIME, SHAP
• Grad-CAM u forenzičkom kontekstu
• Pravni okvir i AI dokazi: EU AI Act, chain of custody, etički izazovi , 2h
• Ishodi: AI u istražnom procesu: integrirana analiza kompleksnih slučajeva
• Prezentacije studentskih projektnih radova

1. Prepare a forensic report suitable for use in legal proceedings in accordance with the EU AI Act.
2. Collect, preserve and analyze digital evidence related to AI systems in accordance with chain of custody principles.
3. Collect, preserve and analyze digital evidence related to AI systems in accordance with chain of custody principles.
4. Apply detection methods for AI-generated text, visual and audio content
5. Describe the architecture and functioning of modern AI systems relevant for forensic analysis (LLMs, diffusion models, GANs)

Required reading

• Tolosana, R. et al., "DeepFakes and Beyond: A Survey of Face Manipulation and Fake Detection", Information Fusion, 2020.
  Casey, E., "Digital Evidence and Computer Crime", Academic Press, 2011.
  Kirchenbauer, J. et al., "A Watermark for Large Language Models", ICML 2023.
  EU Artificial Intelligence Act – Uredba (EU) 2024/1689. 5. NIST AI Risk Management Framework (AI RMF 1.0), 2023.