Security is a major aspect of communication in every day to day life. Neural cryptography is a non-classical cryptographic technique which aims to derive a common key through learning mechanism between two neural networks. It makes use of lightweight permutation parity machines (PPM) to generate a common secret key over a public channel. The basic arithmetic computation on PPM leads to secret key generation between two entities through the process of synchronization. However, the process can be impersonated by an attacker using probabilistic attack on PPM. The proposed technique mitigates the probabilistic attack using feedback mechanism for the synchronization of PPM.