This paper proposes a new approach for detecting and identifying arc flash faults in power systems. The proposed approach is structured to extract the phases of transient frequency components present in arc flash fault currents. The desired phases can be extracted by processing fault currents using a modulated filter bank that is composed of high pass finite impulse response (FIR) filters. These filters are designed by using the Kaiser window method to achieve linear phase responses. Extracting the phases of transient frequency components, present in arc flash fault currents, can provide signature information for accurate and fast detection and distinguish of an arc flash fault. The proposed phase-based approach is implemented for off-line testing to evaluate its performance. Test cases of parallel and series arch flash faults are conducted for supplying linear, non-linear, and dynamic loads. Simulation and off-line results demonstrate the validity, accuracy, speed, and reliability of the phase-based approach to detect and distinguish arc flash faults with minor sensitivities to the load type and arch flash type.