We consider blind decision-feedback equalization (DFE) under high-order modulation, which presents more difficult operating requirements than lower-order modulation. We base our design on the multimodulus algorithm (MMA). To attain fast convergence speed and low steady-state mean-square error (MSE), we consider varying the adaptation step size according to the presently achieved MSE. For this we investigate the properties of the MSE under blind MMA-based DFE and, based on the results, propose a method to estimate its value. The estimate is obtained by analyzing those equalizer filter outputs whose values fall outside the boundary of the modulation's constellation. Simulation results demonstrate the effectiveness of the proposed scheme.