The invention relates to the technical field of
data security, in particular to a defense method for
federated learning depth gradient leakage attacks, which is characterized in that a new joint transformation framework is designed by focusing on gradient vectors, and high precision in
federated learning can be kept while lossless
privacy protection is realized; and meanwhile,
privacy protection, training precision and calculation efficiency are considered through a gradient transformation method, and the original gradient is covered by using the correlation between the gradient and data and adopting gradient difference of two continuous iterations, so that the possibility that an attacker obtains private information by analyzing the uploaded gradient is effectively reduced. In addition, aiming at the limitation that a
homomorphic encryption technology can only be applied to integer calculation, an
encryption scheme is improved, so that the
encryption scheme can support integer components of floating-point numbers, and the limitation of a Paillier
encryption scheme in
processing non-integer data is effectively overcome. According to the method, the high training precision and efficiency of the federal
system are kept, and meanwhile, the
privacy protection effectiveness is improved.