A Game Model for Adversarial Classification in Spam Filtering

Wei Deng; Ze Hui Qu; Li Ye; Zhi Guang Qin

doi:10.4028/www.scientific.net/AMR.433-440.5053

Paper Titles

The Expected Discounted Penalty Function under Stochastic Discount Interest Force Driven by Markov Process
p.5035

Research on Disruption Management for Multi-Objective Problem and its Solution by Genetic Algorithm
p.5040

Design and Implementation of 3D Facial Animation Based on MPEG-4
p.5045

Finite Element Analysis of Woodball Mallet during Swing Stage
p.5050

A Game Model for Adversarial Classification in Spam Filtering
p.5053

Cross-Layer QoSArchitecture for Mobility Management over Heterogeneous Wireless Networks
p.5058

Approximate Packet Delay Analysis for IEEE 802.16 Networks
p.5063

The Design of the Power Saving Mechanisms for IEEE 802.16e Networks
p.5068

The Information Theoretic Perspective on Context Aware Network Design
p.5073

HomeAdvanced Materials ResearchAdvanced Materials Research Vols. 433-440A Game Model for Adversarial Classification in...

A Game Model for Adversarial Classification in Spam Filtering

Abstract:

With the wide applications of machine learning techniques in spam filtering, malicious adversaries have been increasingly launching attacks against the filtering systems. In this paper, we model the interaction between the data miner and the adversary as Stackelberg games. Though existing algorithms for Stackelberg games efficiently find optimal solutions, they critically assume the follower plays optimally and rationally. Unfortunately, in real-world applications, because of follower's bounded rationality and limited observation of the leader's strategy, it may deviate from their expected optimal response. Considering this crucial problem, this paper solve for the Nash equilibrium. Experiments on real spam dataset demonstrate that bounded rationality and limited observation can make Stackelberg games more practical and provide interesting insights about the interaction between the data miner and the spammer in spam filtering.