The cathode material of lithium-ion batteries often contains magnetic foreign materials, which are separated using magnetic separators. However, these materials exhibit weak magnetism, leading to insufficient separation as strong magnetic force separators adsorb them all. Measuring paramagnetic samples is challenging owing to small sample sizes and errors caused by magnetic interference and vibrations. Therefore, a system that can detect weak magnetism in small samples, correct noise, and calculate stable values must be developed. Additionally, foreign materials must be classified to identify their sources effectively. In this study, magnetic field analysis using the finite element method was employed to generate a uniform high-gradient magnetic field for measuring slight weight changes induced by the magnetic field. Additionally, an auxiliary classifier generative adversarial network was utilized to reduce measurement noise in electronic balances, improve accuracy through low-variation clustering, and evaluate differences in magnetic susceptibility caused by the percentage of work-induced martensitic phases in trace paramagnetic materials. This approach enables the efficient identification and separation of foreign material sources.