Encrypted traffic flows have been known to leak information about their underlying content through statistical properties such as packet lengths and timing. While traffic fingerprinting attacks exploit such information leaks and threaten user privacy by disclosing website visits, videos streamed, and user activity on messaging platforms, they can also be helpful in network management and intelligence services. Most recent and best-performing such attacks are based on deep learning models. In this thesis, we identify multiple limitations in the currently available attacks and defenses against them. First, these deep learning models do not provide any insights into their decision-making process. Second, most attacks that have achieved very high accuracies are still limited by unrealistic assumptions that affect their practicality. For example, most attacks assume a closed world setting and focus on traffic classification after event completion. Finally, current state-of-the-art defenses still incur high overheads to provide reasonable privacy, which limits their applicability in real-world applications. In order to address these limitations, we first propose an inline traffic fingerprinting attack based on variable-length sequence modeling to facilitate real-time analytics. Next, we attempt to understand the inner workings of deep learning-based attacks with the dual goals of further improving attacks and designing efficient defenses against such attacks. Then, based on the observations from this analysis, we propose two novel defenses against traffic fingerprinting attacks that provide privacy under more realistic constraints and at lower bandwidth overheads. Finally, we propose a robust framework for open set classification that targets network traffic with this added advantage of being more suitable for deployment in resource-constrained in-network devices.