Classifier evaluation and attribute selection against active adversaries

摘要: Many data mining applications, such as spam filtering and intrusion detection, are faced with active adversaries. In all these the future sets training set no longer from same population, due to transformations employed by Hence a main assumption for existing classification techniques holds initially successful classifiers degrade easily. This becomes game between adversary miner: The modifies its strategy avoid being detected current classifier; miner then updates classifier based on new threats. this paper, we investigate possibility of an equilibrium in seemingly never ending game, where neither party has incentive change. Modifying causes too many false positives little increase true positives; changes decrease utility negative items that not detected. We develop theoretic framework behavior adversarial applications can be analyzed, provide solutions finding point. A classifier's performance indicates eventual success or failure. could select attributes their performance, construct effective classifier. case study online lending demonstrates how apply proposed real application.