Optimal auction mechanism design has long been a focus in computer science and economics. While substantial progress has been made in single-item auctions, optimal design for multi-item auctions has yet to be derived. Recent years have seen a surge in deriving near-optimal auctions through deep learning. As one of the approaches, ALGNet models the bidding process as a two-player game. The ALGNet model however adopted a rather simple design for generating optimal misreports to derive the regret of the trained auction mechanisms. We show that this design can be improved both in network structure and the testing method. Specifically, we train a misreport network tailored for each individual bidder which leads to better misreports. This approach is especially effective when the auctions are asymmetric. By studying misreport, we can get a more accurate estimate of the regret in the auction mechanism thus enhancing its robustness. Experimental results demonstrate that our approach can detect misreport more effectively than previous methods resulting in an increase in regret values as large as 70%. The new misreport network can also be applied to train auction mechanisms, allowing for a better description of the auction process.