Face quality scores can backfire: higher quality often signals a morphing attack
FraunhoferIGD · hf · 2026-11-03
Fraunhofer IGD explores training-free morphing attack detection via face image quality assessment (FIQA), with counterintuitive findings: morphs often score higher quality than bona fide faces, so high quality can indicate failed detection. Performance varies wildly by method and attack type: DifiQA(R) hits 0.0% D-EER on FaceMorpher but 30.0% on DCMorph; FROQ reaches 94.9% on DCMorph but only 31.6% on FaceMorpher; GraFIQs(L) nearly fails on both (49.8%/9.2%). Detection depends heavily on attack type and dataset, showing a single quality signal doesn't generalize.
More from Safety
- AI safety researcher: technical 'AI Safety' work has been product work empowering big labs — j_foerst · 2026-11-03
- CJEU advocate general: chatbots' real-time reuse of press content can infringe publishers' rights — Loo_Atreides · 2026-11-03
- Red teamer: agentic attacks don't need zero-days, they exploit your security debt — AccBalanced · 2026-11-03
- Questions mount as Anthropic reportedly withheld CVE-2026-98432 exploit details from Ubuntu — jedisct1 · 2026-11-03
- Ehud Reiter: four types of LLM-written papers, and why checking matters — EhudReiter · 2026-11-03
- UK AISI spent just £3.34M on compute in 2025-26; researcher says it needs 1000x — HaydnBelfield · 2026-11-03