Independent Reading · 2021
Partial Adversarial Behavior Deception in Security Games
Thanh H. Nguyen, Arunesh Sinha, and He He · Conference paper
Security GamesBehavioral ModellingAdversarial Learning
Full text
Read the Paper
Use View PDF or Download PDF to read the full paper on this device.
Paper Information
- Authors
- Thanh H. Nguyen, Arunesh Sinha, and He He
- Venue
- Conference paper · 2021
Summary
This paper considers a population of attackers in which only part of the observed behaviour is controlled by a deceptive adversary.
Core Idea
Even limited control over training data can distort a learned attacker model and reduce the effectiveness of a defender’s strategy.
Method
The paper introduces a partial behavior deception model and the GAMBO gradient-based algorithm for computing deceptive attack strategies.
Key Takeaways
- A single deceptive participant can influence a model learned from a broader population.
- Partial manipulation can produce meaningful attacker gains and defender losses.
- Security learning systems need safeguards against strategically contaminated data.
Questions
- How much deceptive data is required before a defender’s model becomes unreliable?
- Can robust estimation reduce the impact of partial behavioural manipulation?