Defeating botnet is the key to secure Internet. A lot of cyber attacks are launched by botnets including DDoS, spamming, click frauds and information thefts. Despite of numerous methods have been proposed to detect botnets, botnet detection is still a challenging issue, as adversaries are constantly improving bots to write them stealthier. Existing anomaly-based detection mechanisms, particularly network-based approaches, are not sufficient to defend sophisticated botnets since they are too heavy or generate non-negligible amount of false alarms. As well, tracing attack sources is hardly achieved by existing mechanisms due to the pervasive use of source concealment techniques, such as an IP spoofing and a malicious proxy. In this paper, we propose a host-based mechanism to detect bots at the attack source. We monitor non-human generated attack traffics and trace their corresponding processes. The proposed mechanism effectively detects malicious bots irrespective of their structural characteristics. It can protect networks and system resources by shutting down attack traffics at the attack source. We evaluate our mechanism with eight real-life bot codes that have distinctive architectures, protocols and attack modules. In experimental results, our mechanism effectively detects bot processes in around one second after launching flood attacks or sending spam mails, while no false alarm is generated.