Tor is a widely used low-latency anonymous communication network that hides the IP addresses and communication content of both parties when users are engaged in network activities. Nevertheless, attackers using Website Fingerprinting (WF) attacks can infer the real purpose of websites visited by users through eavesdropping on encrypted network traffic between Tor users and web servers, posting a threat to user privacy and security. Classical WF defense models perform poorly against the latest deep learning-based WF attack models. Furthermore, these models suffer from unreasonable assumptions and high performance overhead. In this paper, we propose a novel zero-latency lightweight WF defense model Multiple Sample Padding (MSP). The model combines multiple traffic sequence features to pad real traffic and defend against deep learning-based website fingerprinting attacks. In addition, we explore the feasibility of changing webpage loading states to interfere with website fingerprint attack models. The experimental results show that MSP can reduce the accuracy of the attack model by 71%, as well as lower the latency overhead by 27%. Changing the webpage loading status based on MSP can reduce the accuracy of the attack model by 90% with a bandwidth overhead of 164%, compared to the strongest WF defense model, while reducing the latency overhead by 78%.