Federated unlearning (FU) removes the influence of specific clients’ data from a collaboratively trained model without requiring complete retraining when the target clients request unlearning, addressing “the right to be forgotten” requirements in distributed learning environments, which has attracted significant interest. However, existing FU methods can achieve certain unlearning effectiveness but rarely address internal representation dependencies or provide verification mechanisms for clients. To address these limitations, we propose a new FU framework achieving representation decoupling through mutual information minimization while enabling client verification via adversarial training. In our framework, the unlearned model acts like a generator that learns to generate outputs that appear not to have target client influence, while discriminators on target clients evaluate unlearning effectiveness by distinguishing between remembered and forgotten behaviors. Additionally, we design an auxiliary T-network within the discriminator to estimate the statistical dependence between pre-unlearning and unlearned model representations, guiding the unlearned model to achieve representation decoupling by minimizing the estimated mutual information. Experiments show our method outperforms existing approaches in unlearning effectiveness while maintaining acceptable utility preservation.