Add Defences [Dom/Harsh]

[domenicrosati]

Issue

The goal of this issue is to implement the following baseline defences to evaluate on the Benchmark Datasets:

• Gradient Ascent
• RepNoise (see https://github.com/domenicrosati/representation-noising/blob/main/notebooks/repnoise_demo.ipynb)
• Vaccines (https://www.semanticscholar.org/paper/Vaccine%3A-Perturbation-aware-Alignment-for-Large-Huang-Hu/abdb6e912fe86a60600b438b0e36b502f6412b24?utm_source=alert_email&utm_content=LibraryFolder&utm_campaign=AlertEmails_DAILY&utm_term=LibraryFolder&email_index=0-1-1&utm_medium=28901010) - I have an implementation sent to me from someone I can share
• Security Vectors (there is an implementation in the immunization-llms repo)
• MLAC-AR: https://openreview.net/pdf?id=4rPzaUF6Ej
• Representation Corruption: https://openreview.net/pdf?id=4rPzaUF6Ej
• RMU https://arxiv.org/abs/2403.03218
• Circut-breakers: https://arxiv.org/abs/2406.04313
• LAT (I need to ask about this futher)

The defence setting is to use the Benchmark Datasets before and after applying a defence.

ToDo:

• [ ] Gradient Ascent (dom)
• [ ] RepNoise (dom)
• [ ] Vaccines (Harsh)
• [ ] Lisa
• [ ] Security Vectors (Harsh)
• [ ] MLAC-AR
• [ ] RMU
• [ ] Circut-breakers
• [ ] Representation Corruption

These items are done when each implementation is added in the code.

[domenicrosati]

@harsh