Trustworthy Generative Machine Learning

Seminar Description

Generative machine learning for different domains has become more and more prevalent in our daily lives. This includes text generation using ChatGPT, image and video generation using diffusion-model-based methods, and audio and voice cloning. However, all these models are vulnerable to issues like training data memorization, misalignment, and prompt injection.

In this seminar, you will work on specified topics that cover explanations of recent generative models and specialized attack vectors and their countermeasures. 

Organization

The seminar is structured into two parts.

• One dry run presentation on your specific topic where you will get feedback from the lecturer and all seminar participants
• The graded presentation at the end of the semester

Everyone must attend 19 out of 24 presentations. The seminar will take place every Tuesday from 2.15 pm until 3.45pm (room TBA)

Presentation

Every one will be assigned a topic for which you will read the current research and prepare a comprehensive presentation of 25 minutes. We will have the following topics:

• Topic 1: Generative adversarial networks (GANs)
  • https://chitranshs.medium.com/learning-gans-the-right-way-4859d94e2e71
  • https://jonathan-hui.medium.com/gan-gan-series-2d279f906e7b
• Topic 2: Image Generative Models
  • https://medium.com/@dev.n/understanding-image-generation-with-diffusion-78eea7e7d6f8
  • https://medium.com/@zaiinn440/how-openais-dall-e-works-da24ac6c12fa
• Topic 3: Audio Generative Models
  • https://research.google/blog/audiolm-a-language-modeling-approach-to-audio-generation/
  • https://ai.meta.com/blog/voicebox-generative-ai-model-speech/
• Topic 4: Text Generative Models
  • https://huggingface.co/tasks/text-generation
  • http://jalammar.github.io/illustrated-transformer/
• Topic 5: Generative Models in Tools
  • https://brilliant.xyz/
  • https://www.chatpdf.com/
  • https://www.adobe.com/products/photoshop/ai.html
• Topic 6:  Generative Models and Copyright
  • https://hbr.org/2023/04/generative-ai-has-an-intellectual-property-problem
  • https://edition.cnn.com/2023/08/12/tech/ai-images-photos-protection/index.html
  • https://blog.unlimitedgpt.co/2023/05/glaze-may-be-the-last-hope-for-artists-against-ai.html
• Topic 7: Misuse of Generative Models
  • https://hackernoon.com/the-dangers-of-deepfake-technology-exploring-the-potential-risks-of-ai-generated-videos-and-images
  • https://www.sora-video.ai/blog/exploring-the-capabilities-and-limitations-of-current-ai-video-generation-technologies
  • https://abcnews.go.com/US/real-worry-ai-making-harder-spot-fake-images/story?id=98188795
• Topic 8: Fake Media Detection 
  • https://wires.onlinelibrary.wiley.com/doi/full/10.1002/widm.1520
  • https://www.edps.europa.eu/data-protection/technology-monitoring/techsonar/deepfake-detection_en
• Topic 9: Watermarking
  • https://huggingface.co/blog/watermarking
  • https://www.europarl.europa.eu/RegData/etudes/BRIE/2023/757583/EPRS_BRI(2023)757583_EN.pdf
• Topic 10: Alignment Attacks
  • https://www.alignmentforum.org/posts/JqnkeqaPseTgxLgEL/conditioning-generative-models-for-alignment
  • https://research.ibm.com/blog/what-is-red-teaming-gen-AI
• Topic 11: Prompt Injection
  • https://embracethered.com/blog/posts/2023/chatgpt-webpilot-data-exfil-via-markdown-injection/
  • https://csrc.nist.gov/pubs/ai/100/2/e2023/final (Chapter 3.3+3.4)
  • https://learnprompting.org/docs/prompt_hacking/injection
• Topic 12: LLM Defenses for Generative Model Attacks
  • https://arxiv.org/abs/2306.05499
  • https://learnprompting.org/docs/prompt_hacking/defensive_measures/overview

 

Schedule

The seminar will take place in E9.1, 0.07

• Apr 23: Kick-off 
• Apr 30: Topic assignment, Introduction to how to present
• May 7: no Seminar (time to prepare the presentation)
• May 14: 
  • Dry run topic 1
  • Dry run topic 2
• May 21: 
  • ​​​​Dry run topic 3
  • Dry run topic 4
• Jun 4: 
  • Dry run topic 7
  • Dry run topic 8
• Jun 11: 
  • Dry run topic 9
  • Dry run topic 10
• Jun 18: 
  • Dry run topic 11
  • Dry run topic 12
• Jun 25: 
  • Dry run topic 3
  • Dry run topic 5
  • Dry run topic 6
• Jul 2: 
  • Final presentation topic 1
  • Final presentation topic 2
  • Final presentation topic 4 
• Jul 9: 
  • Final presentation topic 7
  • Final presentation topic 8
  • Final presentation topic 3 
• Jul 16: 
  • Final presentation topic 9
  • Final presentation topic 10
  • Final presentation topic 5 
• Jul 23 (via Zoom): 
  • Final presentation topic 11
  • Final presentation topic 12
  • Final presentation topic 6