StriderSPD: Gait & Patch Detection Methods

• StriderSPD is a dual-purpose framework defined by innovations in biomedical gait EMG segmentation and binary security patch detection with adaptive processing and multi-modal fusion.
• The gait segmentation pipeline employs synchronized 3-axis IMU and multi-channel sEMG data with adaptive filtering and modality-specific thresholds to achieve high sensitivity and specificity.
• The binary security patch detection component utilizes a dual-branch architecture that fuses GCN-based structural cues with LLM-driven semantic analysis to enhance accuracy and reduce false positives.

StriderSPD refers to two independent methodologies: the Simplified markerless stride detection pipeline for surface EMG segmentation in unconstrained gait analysis ("sMaSDP," also termed "StriderSPD") and the Structure-Guided Joint Representation Learning framework for binary Security Patch Detection. Both approaches share the "StriderSPD" denomination but operate in distinct scientific domains: biomedical gait analysis and binary software vulnerability patch detection.

1. StriderSPD for EMG-Gait Segmentation: Pipeline and Methodological Components

The sMaSDP ("StriderSPD") pipeline is a streamlined algorithm for segmenting surface EMG (sEMG) signals according to gait cycles in unconstrained walking environments using synchronous Inertial Measurement Unit (IMU) data (Aguiar et al., 2023). The protocol consists of:

• Data Acquisition: One 3-axis IMU affixed to the dominant ankle (acquisition rate $F_{\mathrm{KIN}}\approx 240$ Hz) records kinematic data, while simultaneous multi-channel surface EMG ($F_{\mathrm{EMG}}\approx2$ kHz) captures muscle activity.
• Pre-processing and Synchronization: IMU signals are interpolated onto the EMG timebase via linear interpolation and smoothed with a moving average filter of length $2N+1$:

$$\bar{y}(i) = \frac{1}{2N+1}\sum_{n=-N}^{+N}\tilde{y}(i+n)$$

• Modality Segmentation: Movement episodes (duty-cycle) are detected and segmented according to walking modality: level ground walking (LGW), ramp ascend/descend (RA/RD), and staircase ascend/descend (SA/SD).
• Heel-Strike (HS) Detection: Vertical acceleration $a_z$ is bandpass filtered via a pair of zero-phase Butterworth filters (high-pass at 9 Hz, low-pass at 6 Hz, order $n=7$), rectified, and local peaks are extracted using the criterion:

$$a_2'(t_p)=0, \quad a_2''(t_p)<0 \quad \text{and} \quad a_2(t_p) > T = \mu_{a_2} + k \sigma_{a_2}$$

where $k\in[1.5,2.5]$ is a modality-dependent threshold multiplier.

• sEMG Segmentation: EMG segments per gait cycle are extracted with a safety margin $\Delta m$:

$$W_i = [t_{HS,i} - \Delta m,\; t_{HS,i+1} + \Delta m]$$

Segment edges are optionally tapered with a Hann window.

2. StriderSPD for Binary Security Patch Detection: Dual-Branch Joint Representation Framework

In the context of vulnerability mitigation, StriderSPD denotes a two-branch, multi-modal neural architecture for binary Security Patch Detection (SPD) leveraging both assembly code structure and pseudo-code semantics (Li et al., 9 Jan 2026):

• Graph Branch: Assembly code is parsed into Control-Flow Graphs (CFGs). Basic blocks are encoded using UniXcoder; a Gated Graph Convolutional Network (GCN) processes these embeddings:

$$h_i^{(t+1)} = \mathrm{ReLU}( W_A^t h_i^t + \sum_{j\to i} [ \sigma(e_{ij}^t) \circ (W_B^t h_j^t) ] )$$

$$e_{ij}^{(t+1)} = \mathrm{ReLU}(W_U^t e_{ij}^t + W_V^t h_i^t + W_W^t h_j^t)$$

Graph vector $f_g$ is then projected via adapters (gated FFNs) into the LLM's internal query, key, and value embeddings.

• LLM Branch: Pseudo-code representation of pre- and post-patch functions is instruction-tuned on models such as Qwen3-8B using prompts assessing whether a patch resolves a vulnerability.
• Fusion and Cross-Attention: Adapter outputs $h_q$, $h_k$, $h_v$, gated by $g_g$, are fused with LLM hidden states at the token level:

$$Q_\mathrm{fuse} = Q_\mathrm{LLM} + \mathrm{Broadcast}(h_q)$$

$$K_\mathrm{fuse} = [h_k \Vert K_\mathrm{LLM}], \quad V_\mathrm{fuse} = [h_v \Vert V_\mathrm{LLM}]$$

Cross-attention produces the final prediction vector.

• Classification: The fused representation is passed through an MLP and sigmoid to yield $P(\mathrm{security\_patch})$.

3. Parameterization, Training, and Adaptability

For sMaSDP ("StriderSPD"):

Adjustable parameters support heterogeneity in walking types and sensor platforms:

• Cadence Bounds: $C_{\min}=80$ spm, $C_{\max}=120$ spm typical for healthy adult gait; narrower boundaries filter out anomalous events (e.g., foot taps).
• Velocity Gating: Detection suppressed when $|v_z|<v_{\min}$, with $v_{\min}\approx0.1$ m/s.
• Threshold Multiplier $k$: Higher $k$ increases specificity; $k\approx2$ achieves $\approx98\%$ true-positive rate in LGW.

Reparameterization is required for modality shifts (ramps, stairs: cadence $C\approx70$–90 spm), new IMU platforms (sampling rates, noise floors), and changes in EMG sensor latency.

For StriderSPD (Security Patch Detection):

A two-stage training schedule aligns branch capacities:

• Stage 1: Instruction-tune the LLM branch on pseudo-code pairs for semantic vulnerability resolution.
• Stage 2: Freeze LLM; optimize the graph branch and adapters for fusion with LLM representations, using binary cross-entropy loss for patch detection.

StriderSPD generalizes across a wide spectrum of LLMs (Qwen3-8B, Llama-3B, Yi-9B, DeepSeek, LLM4Decompile), with the largest gains observed on Qwen3-8B (+32.8% accuracy, –38.3% FPR).

4. Evaluation, Benchmarking, and Ablation Studies

StriderSPD for EMG-Gait:

Performance metrics on healthy and pathological gait:

Modality	Sensitivity	Specificity	F₁ Score
LGW (healthy)	0.99	0.98	0.985
PD (after tuning)	0.96	0.94	0.95

These figures match or exceed state-of-the-art single-sensor IMU pipelines.

StriderSPD for Binary SPD:

Evaluation on a project- and domain-disjoint binary SPD benchmark, covering five projects and five GCC optimization levels, yields:

Model	Accuracy	F1	FPR
StriderSPD	0.854	0.885	0.293
Yi-9B-Chat	0.758	0.818	0.477

Ablation demonstrates the indispensability of each architectural element; omission of graph branch, adapters, gate, cross-attention, or two-stage training degrades accuracy, F1, or false positive rates by 18–30%.

5. Practical Applications and Adaptation to New Domains

StriderSPD (EMG): Enables cycle-level sEMG segmentation in naturalistic gait studies including ramps, stairs, and curves, with applicability to both healthy and Parkinson’s Disease populations. Easily transferable to new IMU sensors and EMG platforms via parameter re-tuning.

StriderSPD (Binary SPD): Facilitates robust security patch detection in closed-source binaries, overcoming deficits of prior approaches that focused solely on semantic or structural representations. Extensible to multiple LLM architectures and amenable to adaptive thresholding for diverse real-world compilation settings.

6. Contributions and Implications

Both instantiations of StriderSPD offer methodologically transparent frameworks that fuse structural and temporal features via adaptive signal processing (in gait) and architectural integration of modality-specific neural encoders (in binary SPD). The empirical improvements—higher sensitivity, specificity, and F1 in gait; notable advantages in accuracy and reduced false positive rates in patch detection—demonstrate paradigm robustness. The dual-stage optimization in patch detection suggests a general principle for balancing multi-branch neural architectures with disparate parameter sizes. A plausible implication is the utility of adapter-mediated multi-modal fusion for other heterogeneous representation tasks.

7. Common Misconceptions and Clarifications

• "StriderSPD" is not a single algorithm but a term adopted for two distinct methodologies in biomedical signal segmentation (Aguiar et al., 2023) and binary patch detection (Li et al., 9 Jan 2026).
• The EMG segmentation pipeline does not require complex machine learning or labeled gait data, relying instead on signal-processing and adaptive heuristics.
• The binary SPD framework does not depend on source code availability, nor does it solely rely on code semantics; instead, it combines assembly-level structural cues via GCN with high-level pseudo-code features via LLM.

In summary, StriderSPD denotes rigorously evaluated pipelines that advance state-of-the-art cycle-level segmentation in unconstrained gait EMG analysis, and binary security patch detection via structure-guided joint neural representation learning. Both approaches emphasize transparency, adaptability, and superior benchmark performance, making them highly relevant in their respective fields.