This work presents a systematic investigation into the latent knowledge encoded within Network Foundation Models (NFMs). Different from existing efforts, we focus on hidden representations analysis rather than pure downstream task performance and analyze NFMs through a three-part evaluation: Embedding Geometry Analysis to assess representation space utilization, Metric Alignment Assessment to measure correspondence with domain-expert features, and Causal Sensitivity Testing to evaluate robustness to protocol perturbations. Using five diverse network datasets spanning controlled and real-world environments, we evaluate four state-of-the-art NFMs, revealing that they all exhibit significant anisotropy, inconsistent feature sensitivity patterns, an inability to separate the high-level context, payload dependency, and other properties. Our work identifies numerous limitations across all models and demonstrates that addressing them can significantly improve model performance (up to 0.35 increase in $F_1$ scores without architectural changes).

Machine learning is increasingly used in programmable data planes, such as switches [4, 12, 13] and smartNICs [1, 16], to enable real-time traffic analysis and security monitoring at line rate. Decision trees (DTs) are particularly well-suited for these tasks due to their interpretability and compatibility with the Reconfigurable Match-Action Table (RMT) architecture. However, current DT implementations require collecting all features upfront, which limits scalability and accuracy due to constrained data plane resources.

The remarkable success of the use of machine learning-based solutions for network security problems has been impeded by the developed ML models' inability to maintain efficacy when used in different network environments exhibiting different network behaviors. This issue is commonly referred to as the generalizability problem of ML models. The community has recognized the critical role that training datasets play in this context and has developed various techniques to improve dataset curation to overcome this problem. Unfortunately, these methods are generally ill-suited or even counterproductive in the network security domain, where they often result in unrealistic or poor-quality datasets.

Several recent research efforts have proposed Machine Learning (ML)-based solutions that can detect complex patterns in network traffic for a wide range of network security problems. However, without understanding how these black-box models are making their decisions, network operators are reluctant to trust and deploy them in their production settings. One key reason for this reluctance is that these models are prone to the problem of underspecification, defined here as the failure to specify a model in adequate detail. Not unique to the network security domain, this problem manifests itself in ML models that exhibit unexpectedly poor behavior when deployed in real-world settings and has prompted growing interest in developing interpretable ML solutions (e.g., decision trees) for "explaining'' to humans how a given black-box model makes its decisions. However, synthesizing such explainable models that capture a given black-box model's decisions with high fidelity while also being practical (i.e., small enough in size for humans to comprehend) is challenging.