Harnessing effectively the power of cloud computing and the benefits of ubiquitous data collection requires parallel advances in parallel algorithms, also known as multi-agent optimization. Unfortunately, these methods are vulnerable to cyber-attacks: if one or more of the platforms used is compromised and spreads incorrect values to other servers, the answers produced will be incorrect. It is not always possible to prevent this kind of attacks from occurring by relying on authentication alone.
This project studies the vulnerabilities that exist in parallel algorithms, with the intent of understanding how to detect dysfunctional agents, isolate those that are compromised and restore the system functionality. The study of vulnerabilities in decentralized computation has important ramifications that go beyond the engineering discipline. In social science multi-agent optimization is used as a model for social learning. The study of malicious behavior will capture the effect of zealots in social settings that inject false information, steering the outcome of collective decisions towards specific actions that favors their interests.
