The generalization of deepfake detectors to unseen manipulation methods is a critical challenge due to model overfitting on method-specific patterns. While adversarial learning is a common remedy, its conventional error-driven objective is flawed because it only encourages a discriminator to be wrong, which can be achieved without merging the distinct feature clusters of different forgery types. We argue for a paradigm shift from making the discriminator wrong to making it utterly confused. To this end, we propose Maximum Entropy Adversarial (MEA) learning, where the feature extractor is trained to generate features that maximize the entropy of the discriminator’s predictions. This objective, by driving the output distribution towards uniformity, compels the learning of a truly unified, method-agnostic representation by collapsing different forgery features into a single cluster. MEA is the core of a synergistic framework, bolstered by a Supervised Contrastive loss to maintain a clear real-vs-fake margin, and a novel Adversarially-Guided Feature Augmentation (AGFA) that creates challenging boundary samples to strengthen generalization. Extensive experiments on benchmarks like Celeb-DF and DFDC validate our approach, demonstrating our method’s superior generalization.