KAIzen is a zero-knowledge proof of training (zkPoT) for deep neural networks (DNNs) that achieves provable security, succinct proof size, and efficient prover runtime. The system enables a prover to iteratively train a model using mini-batch gradient descent, generating a succinct zkPoT after each iteration to attest to the correctness of the training process. The proof size and verifier runtime are independent of the number of iterations, and the prover runtime is 24× faster than generic recursive proofs, with 27× lower memory overhead. The proof size is 1.63 MB, and the verifier runtime is 130 milliseconds, both independent of the number of iterations and dataset size. KAIzen uses an optimized GKR-style proof system for gradient descent, enabling efficient prover cost and recursive composition of proofs across iterations. The system also introduces an aggregatable polynomial commitment scheme for multivariate polynomials, reducing prover overhead and enabling efficient verification. KAIzen is evaluated on a VGG-11 model with 10 million parameters and batch size 16, achieving 15 minutes per iteration for prover runtime and 1.63 MB proof size. The system is compared to generic IVCs, achieving 24× faster prover time and 27× lower memory usage. KAIzen's construction addresses the limitations of prior work, offering a practical and efficient solution for verifying the training of DNNs.KAIzen is a zero-knowledge proof of training (zkPoT) for deep neural networks (DNNs) that achieves provable security, succinct proof size, and efficient prover runtime. The system enables a prover to iteratively train a model using mini-batch gradient descent, generating a succinct zkPoT after each iteration to attest to the correctness of the training process. The proof size and verifier runtime are independent of the number of iterations, and the prover runtime is 24× faster than generic recursive proofs, with 27× lower memory overhead. The proof size is 1.63 MB, and the verifier runtime is 130 milliseconds, both independent of the number of iterations and dataset size. KAIzen uses an optimized GKR-style proof system for gradient descent, enabling efficient prover cost and recursive composition of proofs across iterations. The system also introduces an aggregatable polynomial commitment scheme for multivariate polynomials, reducing prover overhead and enabling efficient verification. KAIzen is evaluated on a VGG-11 model with 10 million parameters and batch size 16, achieving 15 minutes per iteration for prover runtime and 1.63 MB proof size. The system is compared to generic IVCs, achieving 24× faster prover time and 27× lower memory usage. KAIzen's construction addresses the limitations of prior work, offering a practical and efficient solution for verifying the training of DNNs.