Drone Signal OODD Algorithm

• The paper introduces a drone signal OODD algorithm that detects out-of-distribution signals by identifying change-points in RF and telemetry data.
• It employs methods like time-frequency analysis with DDSCS, multi-modal fusion using Zadoff-Chu sequences, and adaptive feature weighting to enhance detection accuracy.
• RIQN-based temporal dynamics further improve detection delay and robustness, achieving high AUROC and performance across varying noise and protocol conditions.

A drone signal out-of-distribution detection (OODD) algorithm is a system designed to identify whether received drone communications or telemetry signals are statistically inconsistent with patterns observed during training. Such algorithms are essential in drone remote identification (RID), spectrum monitoring, and flight safety because they enable the early detection of new, anomalous, or spoofed drone transmissions, including out-of-band attacks, shifting radio environments, and previously unseen drone protocols.

1. Signal Modalities and OODD Problem Formulations

Drone signals for OODD span two main categories: radio-frequency (RF) I/Q samples and drone telemetry vectors (such as IMU, GPS, barometric, and attitude data). In the OODD context, each signal instance—either a time series $\{x_{1:T}\}$ of $d$-dimensional observations (for telemetry) or raw I/Q samples (for RF)—is assumed to follow an in-distribution model during training ($\mathcal{D}_{\rm train}$). At test time, the core objective is to quickly and reliably detect a change-point $t^*$ such that for $t \ge t^*$, new samples are drawn from an out-of-distribution process $\mathcal{D}_{\rm test} \ne \mathcal{D}_{\rm train}$. The operational requirement is to minimize detection delay while controlling the false positive rate under nominal operation (Danesh et al., 2021).

Different algorithmic approaches have been designed according to the structure of the input signals and the operational constraints; these include time-series dynamics modeling, time-frequency analysis, discriminability-driven attention, and multi-modal feature fusion.

2. Time-Frequency-Based OODD with Discriminability-Driven Spatial-Channel Selection

A leading family of drone RF OODD approaches extracts time-frequency images (TFI) from I/Q sequences using the short-time Fourier transform (STFT),

$$X(t,f) = \Bigl| \sum_{\tau=-\infty}^{+\infty} x[\tau]\,h^*(\tau - t)\,e^{-j2\pi f\,\tau} \Bigr|$$

mapping to an RGB-like tensor $I \in \mathbb{R}^{3 \times H \times W}$ for neural network processing (Feng et al., 26 Jan 2026).

The Discriminability-Driven Spatial-Channel Selection with Gradient Norm (DDSCS) algorithm leverages a convolutional backbone (MobileNetV2), then adaptively weights the extracted feature maps $$\mathbf{F} \in \mathbb{R}^{C_s \times H_s \times W_s}$$ on both spatial and channel axes. The weighting is derived from protocol-specific inter-class similarity and variance calculations, for both spatial positions $(i,j)$ and channels $k$:

• Spatial weights $W^s_{i,j}$ and channel weights $W^c_k$ are computed using formulas involving inter-class cosine similarity and variance, normalized over the spatial grid and channel depth, respectively.
• The final spatial–channel weighted representation is

$$\mathbf{F}_{sc} = \mathbf{F}_s \odot [W^c \otimes 1]$$

followed by global average pooling and a fully-connected classification layer.

To enhance discrimination of OOD samples, the DDSCS method introduces a gradient-norm metric: the $L_2$ norm of the Jacobian of the maximal logit w.r.t. global pooled features. This quantity captures perturbation sensitivity, as OOD samples near the decision boundary exhibit larger gradient norms. The score is linearly fused with a conventional energy-based score,

$$S_{\rm energy} = \log \sum_{j=1}^{N_{\rm cls}} \exp(z_j)$$

to give a final $S_{\rm fused}$, thresholded for OODD.

Key performance characteristics for DDSCS include strong AUROC (95.77%), accuracy (95.18%), robustness across SNRs ($-15$ to $+$15 dB), and improved detection over strong baseline networks (Feng et al., 26 Jan 2026).

3. Multi-Modal Fusion: Cognitive Integration of ZC Sequences and Time-Frequency Images

To address broader protocol diversity and leverage known protocol idiosyncrasies, multi-modal OODD architectures combine generic TFI features with protocol-specific features such as Zadoff-Chu (ZC) sequences (Li et al., 26 Jan 2026).

This Cognitive Fusion approach executes the following pipeline:

• ZC sequence feature extraction: Cross-correlate the I/Q baseband signal with banks of ZC sequences characteristic of known (e.g., DJI) drone protocols, forming a correlation matrix $R$ as input to a dedicated CNN branch.
• Parallel TFI feature extraction: Standard STFT and log-compressed TFI processed by a MobileNetV4-style branch.
• Multi-Modal Feature Interaction (MMFI): Channel- and spatial-wise attention modules interactively refine both modalities via multi-stage operations—concatenation, spatial/channel pooling, convolutional fusion, and attention mechanisms.
• Adaptive Feature Weighting (AFW): Discrimination scores and adaptive masks are computed along spatial ($W_S$) and channel ($W_C$) dimensions (as explicit functions of inter-class similarity and variance).
• Final Decision: The adaptively weighted, fused features are flattened and classified via Softmax over ($P + 1$) classes (the $P$ known drones plus an explicit OOD class). OOD is signaled when $\max_i \mathrm{Softmax}(z_i) \leq \tau$.

Quantitative evaluation on DroneRFa and RFUAV datasets shows the fusion approach yields 1.7% RID and 7.5% OODD accuracy gains compared to the best single-modality baselines, with consistent robustness to SNR degradations, flight distances, and protocol diversity. The ablation studies confirm that both the inclusion of ZC features and the adaptive attentions substantially boost performance (Li et al., 26 Jan 2026).

4. Temporal-Dynamics-Based OODD Using Recurrent Implicit Quantile Networks

For OODD in drone telemetry and continuous time-series, Recurrent Implicit Quantile Networks (RIQN) provide a state-of-the-art probabilistic prediction model (Danesh et al., 2021). RIQN comprises:

• GRU-based history encoder: $h_t = \mathrm{GRU}(x_t,h_{t-1})$.
• Quantile-conditioned embedding: Apply quantile $\tau \sim U(0,1)$ via the fixed feature mapping $\phi(\tau)$.
• Quantile regression network: Compute $z_t^{(m)}=h_t \odot \phi(\tau_m)$, pass through two fully-connected ReLU layers to produce predicted quantiles $\hat{q}_{t+1}^{(m)}$.
• Training loss: Minimize the Huber-quantile loss $\mathcal{L}(\theta)$ over $M$ quantile samples per timestep and all observed trajectories.

Test-time OOD detection is achieved by comparing observed $x_{t}$ to the RIQN-predicted quantile distribution, computing an $L_1$ anomaly score averaged over $M$ quantiles. A sequential CUSUM filter is then applied to smooth $s_t$ and trigger alarms based on a threshold $h$ chosen to yield the desired false positive rate.

Compared with baseline predictors—including non-probabilistic GRU-based networks and random forests—RIQN consistently achieves superior AUROC and detection delay metrics under dynamic OOD scenarios.

5. Comparative Evaluation, Baseline Algorithms, and Metrics

A range of baselines is used to contextualize the performance of advanced OODD methods:

• Vanilla MobileNet(TFI only), Softmax-based, and uncertainty branch for TFI-based OODD (Feng et al., 26 Jan 2026).
• Reconstruction-based (VAE and Fusion-Recon), and channel-only/spatial-only discriminability ablations for hybrid architectures (Li et al., 26 Jan 2026).
• NPN (Non-Probabilistic Network) and Random Forest for time-series approaches (Danesh et al., 2021).

Standardized metrics include:

• AUROC: Area under the ROC curve, assessing threshold-independent separability of ID vs. OOD.
• False Positive Rate (FPR): Fraction of false alarms during nominal (ID) operation.
• Detection Delay: Average lag between OOD injection and alarm.
• Weighted Evaluation Metric (WEM): Mean aggregate of accuracy, recall, F1, and AUROC (Feng et al., 26 Jan 2026).
• Precision, Recall, Accuracy: Defined for frame-level or event-level OODD, as appropriate.

Empirical evaluation under realistic noise (AWGN, SNR sweeps), protocol shifts, flight geometries, and sampling durations uniformly demonstrate that spatial–channel attention, protocol-aware fusion, and perturbation sensitivity significantly improve robustness, detection speed, and adaption to unseen drone types.

6. Implementation Considerations and Computational Aspects

Practical deployments must address data preprocessing, model complexity, and latency:

• Preprocessing: Signal alignment, sampling, noise filtering, normalization, and, if desired, computation of derivative features.
• Hyperparameters: For DDSCS, optimums are $\alpha=0.1$, $\beta=0.2$, $\lambda=0.2$ (Feng et al., 26 Jan 2026); for RIQN, hidden size $H=64$–128, quantile-embedding size $K=64$, learning rate $10^{-3}$ decayed over epochs (Danesh et al., 2021).
• Computation: TFI extraction typically requires $<$10 ms for $L=10^6$ samples; model inference on modern GPUs or embedded NPUs achieves $>$50 fps (Li et al., 26 Jan 2026). Some methods may be too resource-intensive for ultra-constrained edge hardware, making quantization or pruning strategies necessary in those scenarios.
• Online adaptation: For telemetry-based OODD, CUSUM thresholds can be recalibrated during nominal operation to track gradual concept drift; ensembles of RIQNs can enhance detection at ultra-low FPRs.

7. Limitations, Robustness, and Future Directions

Current OODD algorithms exhibit strong performance across a range of flight, noise, and protocol conditions but also have limitations:

• Protocol Dependence: Multi-modal fusion methods leveraging ZC sequences are optimal for DJI-family drones but require fallback to TFI-only approaches on unknown protocols.
• Model Size & Edge Feasibility: Large CNNs and cross-modal attention modules are computationally intensive and may exceed the capabilities of constrained edge devices; this suggests model compression as a future focus.
• Assumptions on OODD Distribution: Training uses only in-distribution data, with no explicit modeling of OOD types, so performance relies on high discriminability of learned features and robust calibration of detection thresholds.

Potential future work includes automated discovery of ZC-like roots in unknown protocols, integration of additional sensing modalities (e.g., angle-of-arrival, CSI), and enhancing fusion architectures for increased interpretability and explainability in OOD decision-making (Li et al., 26 Jan 2026, Feng et al., 26 Jan 2026, Danesh et al., 2021).