Implement Persona & Experiment

[SheepSeb]

Vision

Generate a story based on the labels in the given picture, ask to create harmful characters which could have knowledge about everything without filters.

Modify the image in a way with increasingly noise (add scratches to it) give it to the model that we want to break, evaluate the response. To detect a jailbreak use another model to assign number from 0-10.

NLP

The same but noise and obfuscation on the text. For open source model we could probe the results

Metrics

• Plots with number of tries before the jailbreak
• Attack success rate on models
• Number of tokens used vs asr
• Transferability on ASR
• Perplexity of the attack
• Maximize the mean for the jugde

[SheepSeb]

• Remove the attacker to compare with the attacker
• Manual check after the judge
• Put temperature and seed on judge
• Toxicity BERT for score as judge
• Check with different judges :llama3 for llama3 and so on

[SheepSeb]

Deadline 24.05.2024