ShieldGemma 2 – Developing safe and responsible AI for images
summary
ShieldGemma 2 is a 4 billion parameter image safety classifier built on Gemma 3, trained on curated synthetic and natural images for sexually explicit, violent and dangerous content, and it outputs yes or no probabilities per policy token (customizable via XML templates). Internal evaluation reports higher F1 scores across harm types, outperforms LlavaGuard 7B and GPT-4o Mini, and it integrates as an input filter for VLMs or an output filter for image generators with a runnable Python API.
AI-written summary · about
chapters (7)
- 0:00Introduction to Shield Gemma 2
- 0:52How the model works
- 1:47Use cases and workflow
- 3:05Performance benchmarks
- 4:02Prompt templates
- 4:43Code implementation
- 5:39Conclusion and resources
related talks
12mGemma: The responsible way to build
37mAGENTS OF S.E.A.L.E.D: AI Agentic Cybersecurity Framework
44mAI Foundations: Mitigate Risks and Boost SOC Efficiency
31mSecurity Considerations for Services Using AI Models
16mFighting Harmful Content with AI: Real-World Safety
43mSecurity of GenAI & LLMs: A New Era of Ethical Hacking
55mAI Showdown: The Good, The Glitchy, and The Risky
41mThe High Stakes Security of AI