High-stakes lying causes detectable changes in human behavior and physiology. Lie detection techniques based on behavior analysis are unobtrusive, but often require laborintensive efforts. Lie detection techniques based on physiological measurements are more amenable to automated analysis and perhaps more objective, but their often obtrusive nature makes them less suitable for realistic studies. In this paper we present a novel lie detection framework. At the core of this framework is a physiological measurement method that quantifies stress-induced facial perspiration via thermal imagery. The method uses a wavelet-based signal processing algorithm to construct a feature vector of dominant perinasal perspiration frequencies. Then, pattern recognition algorithms classify the subjects into deceptive or truthful by comparing the extracted features between the hard and easy questioning segments of an interview procedure. We tested the framework on thermal clips of 40 subjects who underwent interview for a mock crime. We used 25 subjects to train the classifiers and 15 subjects for testing. The method achieved 80% success rate in blind predictions. This framework can be generalized across experimental designs, as the classifiers do not depend on the number or order of interview questions.