Optimal auction mechanism design has long been a focus in computer science and economics. While progress has been made in single-item auctions, optimal design for multi-item auctions is challenging. Deep learning has shown promise in deriving near-optimal auctions, but limited attention has been given to regret truthfulness and identifying near-optimal misreports. Current methods primarily rely on inner loop optimization to detect misreports, which may not guarantee effective enough results and lead to inaccurate regret values. In this study, we introduce the concept of a ‘Multi-Misreport network’ that prioritizes identifying improved misreports to achieve regret values closer to the true ones. Specifically, we train bidder-specific misreport networks, leading to improved misreports. Based on that, a general misreport testing method is proposed to be used as a test plug-in for various mechanisms. Then, we propose a new mechanism training framework that models the auction process as a game between the mechanism and bidders, which we call MAGNet. We show that this design improves both the auction mechanism testing methods and its training methods. Experimental results demonstrate the effectiveness of our approach in detecting superior misreports and obtaining better mechanisms compared to state-of-the-art mechanism designing methods.