Neural cryptography approaches probabilistic algorithm for exchange of keys with assured security. Artificial neural networks do not require complex number theory instead they work on basic arithmetic computation. Tree parity machine (TPM), one of the multi-layer neural network helps in generation of secret key between two ends by the process of synchronization. However the process can be mimicked by an attacker with the powerful majority flipping attack on TPM. The proposed technique mitigates Majority Flipping Attack (MFA) using link weights driven dynamic synchronization of TPM.