This paper investigates effect of malicious intelligence reflecting surface (IRS). The malicious IRS is utilized for performing attack by randomly reflecting data sequences of legitimate users (LUs) to a base station (BS). We find that the data sequences of LUs are correlative to the signals reflected by malicious IRS. The correlation undermines the performance of traditional eigenvalue decomposition (EVD)-based channel estimation (CE) methods. To address this challenge, we propose a empirical-distribution-based channel estimation approach in the presence of malicious IRS. The proposed method works by capturing desired convex hulls from signals disturbed by malicious IRS, on the basis of its empirical distribution. Simulation results show that our proposed approach outperforms traditional EVD-based methods as much as nearly 5 dB in normalized mean square error (NMSE).