In recent years, network security of cyber-physical systems has become a hot research topic. Investigating the problem of designing attacks from the attacker's perspective can effectively evaluate the vulnerability of the system to network attacks, and provide theoretical basis for designing network protection methods. For this reason, this paper investigates the problem of designing the optimal $\epsilon$-stealthy deception attacks against remote state estimation in cyber-physical systems. Firstly, different from the related results which require extra filters and historical data to calculate the true innovation online, this paper proposes a self-generated attack model which uses off-line generated attack signals to tamper with the sensor measurements and deteriorate the estimation performance, such that the attacks are more easily to be implemented. Subsequently, the remote estimation error under the attack is derived to quantify the attack effect, based on which, the attack design problem is transformed into a variable optimization problem. Since the model uses the more general time-varying mean, the optimization problem contains more decision variables, which cannot be solved directly by the attack optimization methods in the related results. Therefore, the problem is equivalently transformed by using the relevant statistical properties of K-L divergence and mutual information. Furthermore, by combining the Lagrange multiplier method and the optimization method with the covering by the related parameter characteristics, the optimal attack strategy is obtained to maximize the remote estimation error under the $\epsilon$-stealthiness. Finally, simulation examples are given to verify the validity of the results.