In order to allocate limited network resources in cloud computing efficiently to ensure Quality of Service （QoS） while improving resource utilization and management efficiency， a security and reliability driven Encoder-Decoder-based Deep Reinforcement Learning （ED-DRL） method was proposed for Service Function Chain （SFC） deployment. In the method， the SFC deployment was regarded as a Markov Decision Process （MDP）， a Graph ATtention network （GAT） encoder and a Gated Recurrent Unit （GRU） decoder were employed to extract network topology features and inter-node dependencies effectively， and an Asynchronous Advantage Actor-Critic （A3C） algorithm was combined to generate SFC deployment strategies dynamically. To address security and reliability requirements， the reward function was redesigned to guide the policy network in selecting optimal resources. Simulation results demonstrate that ED-DRL achieves an acceptance rate of 70.7% and an average revenue of 0.063 5， outperforming comparison methods such as Continuous-Decision scheme relying on Reinforcement Learning （CDRL）.