Chaotic optical secure communication can provide physical-layer security for optical communication. In this paper, we propose an optical chaos shift keying communication scheme based on a deep neural network (DNN) classifier. At the transmitter, digital information is loaded by switching the system's time-delay parameter. At the receiver, a DNN classifier is utilized to achieve demodulation, identifying the trajectory variations caused by switching the time-delay parameter. This approach eliminates the need for carrier recovery, reducing system implementation complexity while enhancing noise immunity. The simulation results demonstrate that at a signal-to-noise ratio (SNR) of 15 dB, the system can support secure communication at a transmission rate of 2.5 Gbps. When the SNR decreases to 10 dB, the transmission rate can be maintained on the order of 1.6 Gbps, and even when the SNR drops to 5 dB, reliable communication is still achievable. Performance comparison analysis shows that, compared with other existing NN-based CSK variants, the proposed method exhibits better anti-interference ability, lower bit error rate (BER), and higher bit transmission rate.