Episode-Based Training

Each episode is constructed by randomly sampling a task from a broader task distribution. A task is defined by a small set of classes (e.g., 5-way) with only a few examples each (e.g., 1-shot or 5-shot). This mirrors the deployment scenario where the model encounters novel devices with minimal enrollment samples.

• N-way, K-shot: Standard notation where N is the number of classes per episode and K is the number of labeled examples per class
• Tasks are sampled across different transmitters, modulation schemes, or environmental conditions
• Prevents overfitting to any single class distribution

Within each episode, the sampled data is partitioned into a support set and a query set. The support set provides the labeled examples used to condition the model, while the query set contains unlabeled samples used to evaluate adaptation performance.

• Support set: The few labeled examples the model uses to infer class boundaries (e.g., 5 samples per class)
• Query set: Held-out examples from the same classes used to compute the loss and update meta-parameters
• This split ensures the model is evaluated on its ability to generalize, not memorize

Episode-based training employs a bilevel optimization structure. The inner loop performs task-specific adaptation using the support set, while the outer loop updates the model's initial parameters across many episodes to learn a generalizable starting point.

• Inner loop: Fast adaptation on the support set, often with only a few gradient steps
• Outer loop: Meta-optimization across episodes to minimize query set error
• This structure is the foundation of algorithms like MAML and Reptile

In Prototypical Networks and similar architectures, each episode computes a prototype for every class by averaging the support set embeddings. Query samples are then classified based on their distance to these prototypes in the learned embedding space.

• Class prototype: The mean vector of support embeddings for a given class
• Distance metrics commonly used include Euclidean distance and cosine similarity
• The entire episode is trained end-to-end to minimize classification error on the query set

By training across thousands of randomly composed episodes, the model experiences a vast array of class combinations and data scarcity scenarios. This acts as a form of implicit cross-validation, ensuring robust performance on truly novel tasks at deployment.

• Each episode is effectively a unique mini-dataset
• The model never sees the same class combination twice during meta-training
• Meta-testing uses entirely held-out classes to evaluate true few-shot generalization capability

In few-shot device enrollment, each episode simulates the onboarding of a small batch of new transmitters. The support set contains the minimal enrollment captures (e.g., 1-5 IQ samples), and the query set tests whether the model can authenticate subsequent transmissions from those same devices.

• Episodes train the model to extract hardware impairment signatures from minimal data
• Directly addresses the cold-start problem in IoT and supply chain authentication
• Enables one-shot enrollment where a single transmission suffices for future identification

A twin-network architecture that learns a similarity metric by comparing pairs of inputs. During an episode, the network processes two IQ samples and outputs a distance score indicating whether they originate from the same transmitter. Training involves minimizing contrastive loss on paired examples sampled from the support and query sets. This verification-based approach is particularly robust for one-shot enrollment scenarios where only a single clean recording of a device is available.

A contrastive loss function that structures the embedding space by enforcing a margin-based relationship between triplets of samples. Each triplet consists of an anchor, a positive sample from the same device, and a negative sample from a different device. The loss penalizes the network if the distance between the anchor and negative is not greater than the anchor-positive distance by a specified margin. Episode-based sampling of hard triplets is critical for learning discriminative RF fingerprints.

The two fundamental data partitions within a single training episode. The support set provides the labeled examples used to condition the model for a specific task—typically 1 to 5 samples per device. The query set contains unlabeled samples used to evaluate the model's generalization performance on that task. This structure explicitly mimics the deployment scenario where a new device is enrolled with minimal captures and must then be authenticated against future transmissions.

A lower-dimensional, continuous vector space where raw IQ samples are mapped by a neural network. The geometry of this space is the direct product of episode-based training. A well-trained embedding function places signals from the same transmitter hardware in tight, separable clusters, while signals from different devices are pushed apart. Distance metrics like cosine similarity or Euclidean distance in this space become the basis for authentication decisions.