An output‐only multiple‐crack localization method is proposed in this paper to detect and localize breathing cracks in a stepped rotor, which utilizes the crack‐induced local shape distortions in super‐harmonic characteristic deflection shapes (SCDSs). To minimize the noise effects on SCDSs and improve the accuracy of SCDS‐based crack localization, singular value decomposition is adopted to estimate the SCDS as the dominant singular vector of output power spectral density matrix at a super‐harmonic frequency. Then, in order to better reveal shape distortions in the SCDSs, an after‐treatment technique called gapped smoothing method is applied to derive a damage index. Numerical experiments are carried out to investigate the performance of the proposed method based on a two‐disc stepped rotor‐bearing system with breathing cracks established by the finite element method. Results show that the method is effective for single and multiple crack localization in stepped rotors and interference of steps can be excluded. Furthermore, the method is robust to noise. Influences of crack depths and rotating speeds are also investigated, and how to choose the rotating speed for better crack localization is discussed.