DARPA Guaranteeing AI Robustness against Deception (GARD) TA2. Scalable and Composable Adversarial Robustness Framework (SCARF) Project
Abstract
The IBM team had the following objectives during the DARPA GARD program: Maintain the Adversarial Robustness Toolbox (ART) to ensure its architecture and all contributions are properly included following best software practices. Extend ART according to the evaluation requirements outlined by DARPA and MITRE. Survey the literature to determine what new techniques arise during the program to maintain a state-of-the art selection of tools for the GARD evaluation team and other performers. Collaborate with TA1 and TA2 performers to enable the addition of their new techniques into ART. Help during the evaluation cycles by assessing some performers submissions. During the program, the expertise of the IBM team in the poisoning area resulted in IBM taking the role of the poisoning team lead. In this role, the IBM team closely collaborated with the other TA2 teams MITRE and Two Six Technologies to evaluate TA1 submissions and helped define and implement scenarios to be used during the evaluation cycles. Nathalie Baracaldo also presented the poisoning team results at the Principal Investigator meetings.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 16, 2024
- Accession Number
- AD1221696
Entities
Organizations
- International Business Machines Corporation (Armonk, NY)