TSLT-Net: Lightweight Temporal-Spatial Transformer IDS

• The paper presents TSLT-Net, a unified lightweight temporal-spatial transformer IDS for drone networks that achieves 99.99% multiclass and 100% binary detection accuracy.
• It uses a shared attention mechanism to capture both sequential temporal patterns and inter-feature spatial dependencies within a single architecture.
• Its compact design, with only 9,722 parameters and a 0.04 MB memory footprint, enables real-time deployment on resource-constrained UAV systems.

Searching arXiv for the specified paper to verify bibliographic details and grounding. Retrieving the arXiv record and a related query for transformer-based drone intrusion detection. TSLT-Net is a lightweight and unified Temporal Spatial Transformer based intrusion detection system tailored specifically for drone networks. Introduced in "A Novel Unified Lightweight Temporal-Spatial Transformer Approach for Intrusion Detection in Drone Networks" (Biswas et al., 3 Oct 2025), it is designed to capture both sequential temporal patterns and inter-feature spatial dependencies in network traffic through self attention, while supporting both multiclass attack classification and binary anomaly detection within a single architecture. On the ISOT Drone Anomaly Detection Dataset, comprising more than 2.3 million labeled records, the reported performance is 99.99 percent accuracy in multiclass detection and 100 percent in binary anomaly detection, with a model size of 0.04 MB and 9 722 trainable parameters (Biswas et al., 3 Oct 2025).

1. Design objectives and problem setting

TSLT-Net is motivated by the cybersecurity demands created by the growing integration of drones across commercial, industrial, and civilian domains. The paper situates the model against the susceptibility of drone networks to a wide range of cyberattacks and argues that existing intrusion detection mechanisms often lack the adaptability, efficiency, and generalizability required for dynamic and resource constrained environments (Biswas et al., 3 Oct 2025).

The stated design goals are threefold. First, the model is lightweight, with only 9 722 trainable parameters and a 0.04 MB model size. Second, it provides unified temporal-spatial modeling, capturing both sequential patterns and inter-feature dependencies in raw packet-flow features. Third, it targets edge suitability through minimal memory footprint and FLOPs, enabling real-time inference for UAV on-board systems (Biswas et al., 3 Oct 2025).

The architecture is explicitly framed as a single model that can serve both a 10-way multiclass task and a 2-way binary task. This unification is central to the formulation: the preprocessing pipeline, feature encoding, attention block, pooling stage, and dense classification pathway are shared, with only the output dimensionality changing between the two settings (Biswas et al., 3 Oct 2025).

2. Input processing and feature representation

The preprocessing pipeline begins with raw PCAP conversion to CSV via CIC-IoT scripts. Feature extraction combines conventional flow features with drone-specific attributes, including Drone_port, DS_status, Entropy, and Payload_Length. Missing values are handled by median imputation for numerical variables and mode imputation for categorical variables (Biswas et al., 3 Oct 2025).

Numerical features are standardized by z-score normalization,

$x_{\text{norm}} = \frac{x - \mu}{\sigma},$

while categorical features are one-hot encoded for TSLT-Net. The description distinguishes this from the treatment used for other models, where categorical features are handled with LabelEncoder (Biswas et al., 3 Oct 2025). The train/test split is 80/20 and stratified.

Figure 1, as described in the paper, organizes the full data path into preprocessing and feature paths on the left, a central sequence-processing block, and a classification tail on the right. The left side specifically combines numerical and categorical paths before concatenation, indicating that the model is intended to ingest heterogeneous network-traffic descriptors rather than a single homogeneous signal representation (Biswas et al., 3 Oct 2025).

3. Unified temporal-spatial transformer architecture

The high-level architecture is summarized as: preprocessing and feature encoding, a dense projection layer, sequence reshape, LayerNormalization, Multi-Head Self-Attention, GlobalAveragePooling, a dense layer, dropout, and a final softmax head (Biswas et al., 3 Oct 2025).

More concretely, the input vector $x \in \mathbb{R}^{d'}$ is first mapped through a dense layer with 128 units and ReLU activation:

$$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$

The resulting representation $h^1 \in \mathbb{R}^{128}$ is reshaped into $H^2 \in \mathbb{R}^{T \times d}$ with $T = 16$ and $d = 8$. LayerNormalization is then applied over each of the $T$ positions. The Multi-Head Self-Attention block outputs a representation in the same shape, $H^4 \in \mathbb{R}^{T \times d}$. This is reduced by GlobalAveragePooling,

$$h^5 = \frac{1}{T}\sum_{i=1}^{T} H^4_{i,\cdot} \in \mathbb{R}^{d},$$

followed by a second dense layer with 64 units and ReLU activation,

$x \in \mathbb{R}^{d'}$0

A dropout layer with rate 0.3 produces $x \in \mathbb{R}^{d'}$1, which is passed to the final dense softmax classifier (Biswas et al., 3 Oct 2025).

A notable design choice is that no explicit residual connections or positional encodings are used in the lightweight design; the paper states that LayerNorm plus MHA suffices for stability. This directly counters a common assumption that transformer-based models must always retain the full residual-and-positional-encoding pattern to remain effective. In TSLT-Net, the omission is a deliberate compression-oriented choice rather than an incidental simplification (Biswas et al., 3 Oct 2025).

4. Attention mechanism and temporal-spatial coupling

The core Transformer block operates on a normalized input $x \in \mathbb{R}^{d'}$2 with $x \in \mathbb{R}^{d'}$3 and $x \in \mathbb{R}^{d'}$4. Each of the $x \in \mathbb{R}^{d'}$5 heads computes queries, keys, and values as

$x \in \mathbb{R}^{d'}$6

with $x \in \mathbb{R}^{d'}$7 and $x \in \mathbb{R}^{d'}$8 (Biswas et al., 3 Oct 2025).

The per-head scaled dot-product attention is

$x \in \mathbb{R}^{d'}$9

The multi-head output is then formed by concatenation and linear projection:

$$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$0

$$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$1

Although implemented as a single MHA block, the two heads are described as jointly learning temporal dependencies across the 16 time positions and spatial correlations among the 8 feature dimensions. The same attention equations apply in both cases; the distinction arises from what each head’s learned linear projections emphasize. This suggests that the temporal-spatial unification is not achieved by separate specialized modules, but by representational factorization within a shared attention operator (Biswas et al., 3 Oct 2025).

5. Classification heads, optimization, and reported performance

The model uses the same core architecture for two output regimes. For multiclass intrusion detection, the final layer produces $$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$2 with softmax activation. For binary anomaly detection, the same design is used with $$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$3. In both cases, the loss is categorical cross-entropy,

$$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$4

The training configuration specifies Adam with learning rate $$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$5, batch size 128, up to 50 epochs with early stopping of patience 5, two transformer heads, $$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$6, and embedding dimensions following the sequence 128 $$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$7 reshape to $$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$8 $$h^1 = \operatorname{ReLU}(W^1 x + b^1), \qquad W^1 \in \mathbb{R}^{128 \times d'}.$$9 MHA $h^1 \in \mathbb{R}^{128}$0 64 $h^1 \in \mathbb{R}^{128}$1 output (Biswas et al., 3 Oct 2025).

The experimental comparison reported on the ISOT Drone Dataset is summarized below.

Model	Accuracy (%)	Footprint
CNN	99.78	0.51 MB / 134 026 params
MLP	99.72	0.76 MB / 199 306 params
GRU	99.79	0.46 MB / 120 970 params
RNN	99.78	0.15 MB / 38 986 params
LSTM	99.79	0.48 MB / 125 962 params
TSLT-Net	99.99	0.04 MB / 9 722 params

For TSLT-Net specifically, the paper reports precision, recall, and F1-score of 99.99 in the deep-learning model comparison, alongside 99.99 percent multiclass accuracy and 100 percent binary anomaly detection accuracy (Biswas et al., 3 Oct 2025). Within the reported benchmark, the model combines the highest listed accuracy with the smallest listed model size and parameter count.

6. Deployment profile, interpretation, and research context

The paper characterizes TSLT-Net as an effective and scalable solution for real time drone cybersecurity, particularly suitable for deployment on edge devices in mission critical UAV systems (Biswas et al., 3 Oct 2025). That assessment is tied directly to the reported memory footprint, parameter budget, and on-board inference orientation rather than to a large or highly layered transformer stack.

The system’s significance lies in the combination of three properties that are often treated separately in intrusion-detection architectures for constrained platforms: unified handling of temporal and spatial dependencies, a single architecture for multiclass and binary settings, and a lightweight implementation compatible with UAV edge deployment. The description repeatedly emphasizes that the model operates on raw packet-flow features after a streamlined preprocessing pipeline, rather than depending on a heavier multi-stage ensemble or a separate temporal module and feature-correlation module (Biswas et al., 3 Oct 2025).

A common misconception would be to interpret TSLT-Net as a generic transformer transplanted into a drone-security setting. The paper instead presents it as a domain-specific intrusion detection design for drone networks, with drone-specific feature extraction and explicit attention to UAV resource constraints. Another possible misconception would be to assume that the model’s temporal-spatial characterization implies two disjoint attention subsystems. The implementation described in the paper uses a single MHA block whose two heads jointly learn both forms of dependency (Biswas et al., 3 Oct 2025).

The paper concludes that this workflow—from raw traffic encoding, through unified temporal-spatial attention, to lightweight inference—demonstrates why TSLT-Net achieves state-of-the-art intrusion detection in drone networks (Biswas et al., 3 Oct 2025). A plausible implication is that the model is intended less as a general-purpose transformer variant than as a compact systems-level design point for drone-network IDS under strict memory and latency constraints.