The escalating sophistication of cyber-attacks and the widespread utilization of stealth tactics have led to significant security threats globally. Nevertheless, the existing static detection methods exhibit limited coverage, and traditional dynamic monitoring approaches encounter challenges in bypassing evasion techniques. Thus, it has become imperative to implement nuanced and dynamic analysis to achieve precise behavior detection in real time. There are two pressing concerns associated with current dynamic malware behavior detection solutions. Firstly, the collection and processing of data entail a significant amount of overhead, making it challenging to be employed for real-time detection on the end host. Secondly, these approaches tend to treat malware as a singular entity, thereby overlooking varied behaviors within one instance. To fill these gaps, we propose PARIS, an adaptive trace fetching, lightweight, real-time malicious behavior detection system. Specifically, we monitor malicious behavior with Event Tracing for Windows (ETW) and learn to selectively collect maliciousness-related APIs or call stacks, significantly reducing the data collection overhead. As a result, we can monitor a wider range of APIs and detect more intricate attack behavior. We implemented a prototype of PARIS and evaluated the system overhead, the accuracy of comparative behavior recognition, and the impact of different models and parameters. The result demonstrates that PARIS can reduce over 98.8% of data compared to the raw ETW trace and hence decreases the overhead on the host in terms of memory, bandwidth, and CPU usage with a similar detection accuracy to the baselines that suffer from the high overhead. Furthermore, a breakdown evaluation shows that 80% of the memory and bandwidth savings and a complete reduction in CPU usage can be attributed to our adaptive trace-fetching collector.