Socio-technical Defense Against Voice Spamming

Voice over IP (VoIP) is a key enabling technology for migration of circuit-switched PSTN (Public Switched Telephone Network) architectures to packet-based networks. One problem of the present VoIP networks is filtering spam calls referred to as SPIT (Spam over Internet Telephony). Unlike spam in e-mail systems, VoIP spam calls have to be identified in real time. Many of the techniques devised for e-mail spam detection rely upon content analysis, and in the case of VoIP, it is too late to analyze the content (voice) as the user would have already attended the call. Therefore, the real challenge is to block a spam call before the telephone rings. In addition, we believe it is imperative that spam filters integrate human behavioral aspects to gauge the legitimacy of voice calls. We know that, when it comes to receiving or rejecting a voice call, people use the social meaning of trust, reputation, friendship of the calling party and their own mood. In this article, we describe a multi-stage, adaptive spam filter based on presence (location, mood, time), trust, and reputation to detect spam in voice calls. In particular, we describe a closed-loop feedback control between different stages to decide whether an incoming call is spam. We further propose formalism for voice-specific trust and reputation analysis. We base this formal model on a human intuitive behavior for detecting spam based on the called party's direct and indirect relationships with the calling party. No VoIP corpus is available for testing the detection mechanism. Therefore, for verifying the detection accuracy, we used a laboratory setup of several soft-phones, real IP phones and a commercial-grade proxy server that receives and processes incoming calls. We experimentally validated the proposed filtering mechanisms by simulating spam calls and measured the filter's accuracy by applying the trust and reputation formalism. We observed that, while the filter blocks a second spam call from a spammer calling from the same end IP host and domain, the filter needs only a maximum of three calls—even in the case when spammer moves to a new host and domain. Finally, we present a detailed sensitivity analysis for examining the influence of parameters such as spam volume and network size on the filter's accuracy.