This paper proposes a tensor-based modulation classification scheme for Multiple Input Multiple Output systems operating under malicious interference. The method begins by constructing a high-dimensional mutual cumulative tensor and applying a natural gradient Candecomp/Parafac (CP) decomposition algorithm for blind interference separation. To reduce complexity, a compact three-dimensional tensor is then adaptively built using time-frequency features. A novel generalized tensor kernel discriminant analysis is introduced for effective feature extraction, combining the structural benefits of tensors with the nonlinear capability of kernel methods. Finally, a -nearest neighbor classifier is used for recognition. Requiring no prior channel or noise information, the proposed scheme demonstrates high classification accuracy in low signal-to-interference ratio environments.