Utility companies consistently suffer from the harassing of Non-Technical Loss (NTL) frauds globally. In the traditional power grid, electricity theft is the main form of NTL frauds. In Smart Grid, smart meters thwart electricity theft in some ways but cause more problems, e.g., intrusions, hacking, and malicious manipulation. Various detectors have been proposed to detect NTL frauds, but they either rely on user behavior analysis which requires a large amount of historical data or needs a lot of extra devices which are expensive. In this paper, a detector named NFD (NTL Fraud Detection), is proposed to detect NTL frauds with only a small amount of data and one additional device. NFD is based on Lagrange polynomial interpolation to model an adversary's behaviors, and detect a tampered meter by comparing the difference between the results. Different from existing detectors, our detector knows adversaries better than adversaries themselves. By building mathematical models of these adversaries, we can predict their behaviors which they may not be aware of by themselves. NFD makes it practical to detect NTL frauds both online and offline in Smart Grid. NFD can facilitate real-world applications before Smart Grid is fully deployed since it can serve the traditional power grid and Smart Grid simultaneously. Experimental results show the effectiveness of NFD. It can detect multiple tampered meters and multiple adversaries, as well as a single tampered meter and a single adversary. We also study how to tune the parameters used in NFD to further guide its practical usage.