This article investigates the design of an adaptive event-triggering control under limited communication resources and deception attacks, ensuring the security and stability of a class of cyber-physical systems with stochastic disturbances and non-affine terms. First, unlike the most existing results, the differentiable requirement imposed on the non-affine terms is substantially extended. A semi-bounded condition is proposed to obtain a pre-affine systems by a model transformation method with a scaling function. Second, a neural network-based learning approximation algorithm is introduced to eliminate the coupling effects between errors and stochastic terms. Third, an event-triggering method is designed to minimize the waste of communication resources. With the designed control input and adaptive laws, all closed-loop signals are semi-globally uniformly ultimately bounded in probability. In the end, an one-link manipulator actuated by a DC motor is exploited to validate the effectiveness of the proposed method.