KEEP TRACE β Public Security Benchmarks (Ameforge)
KEEP TRACE is a localized forensic integrity and threat-analysis engine built on the proprietary KEEP (Keep what Endures under Environmental Perturbation) architecture by Ameforge (Windsor, Ontario).
- Product page: https://ameforge.tech/products/keep-trace
- Key properties: Zero self-attention ($Q K^\top$), zero CNN, zero tokenizer, linear $O(N \cdot k)$ complexity, sub-megabyte footprint, byte-level anomaly localization + supervised PE ransomware classification.
(Note: This Hugging Face repository publishes reproducible benchmark results and methodology only. Proprietary model weights and core architecture remain private on Ameforge servers.)
Head-to-Head Public Benchmarks (NVIDIA RTX 3050, CUDA 12.6)
1. Windows PE Ransomware Detection (Raw 1,024-byte PE Headers)
Dataset: cycloevan/Ransomware_PE_Header_Feature_Dataset (2,157 real Windows PE executables: goodware vs. 25 ransomware families, 80/20 stratified split).
| Model | Size (int8 eq.) | ROC AUC | Accuracy | F1 | Inference |
|---|---|---|---|---|---|
| KEEP TRACE Classifier (0 attn, 0 CNN) | 254 KB | 0.9988 (1st) | 98.1% | 0.980 | 65.5 ms |
| MalConv (Gated CNN baseline) | 74 KB | 0.9982 | 99.1% | 0.990 | 5.5 ms |
| Transformer Encoder (3L, 4H) | 605 KB | 0.9905 | 97.5% | 0.973 | 129.8 ms |
| KEEP TRACE Unsupervised (0 malware seen) | 177 KB | 0.5537 | 61.1% | 0.472 | 53.3 ms |
2. Cross-Distribution PE Malware (mjbommar/binary-30k)
Dataset: mjbommar/binary-30k (3,577 real PE binaries: 2,456 train / 1,121 test with strong class-balance shift).
| Model | Size (int8 eq.) | ROC AUC | Accuracy | F1 | Inference |
|---|---|---|---|---|---|
| CNN 1D (MalConv-style) | 102 KB | 0.9870 | 94.9% | 0.932 | 96.9 ms |
| KEEP TRACE (0 attn, 0 CNN) | 249 KB | 0.9493 | 88.3% | 0.857 | 301.0 ms |
| Transformer Encoder (3L, 4H) | 599 KB | 0.9252 | 80.6% | 0.773 | 110.6 ms |
3. EMBER-2018 Tabular Benchmark (cw1521/ember2018-malware)
Dataset: cw1521/ember2018-malware (16,000 samples: 12,800 train / 3,200 test, 2,381 tabular PE features).
| Model | Size (int8 eq.) | ROC AUC | Accuracy | F1 | Inference |
|---|---|---|---|---|---|
| HistGBDT (LightGBM-equivalent) | GBDT | 0.9948 | 96.7% | 0.965 | 457.0 ms |
| Tabular MLP (256β128) | 629 KB | 0.9824 | 94.5% | 0.942 | 6.3 ms |
| KEEP TRACE (0 attn) | 255 KB | 0.9686 | 91.0% | 0.905 | 223.4 ms |
4. Zero-Signature Localized Injection & Exploit Detection (Native x86-64 .text & Compiled Bytecode)
Evaluated on unseen real compiled .dll/.pyd binaries and real compiled functions with zero attack labels:
- Encrypted / Packed Payload Injection (x86-64
.text): AUC 0.9986 - NOP-Sled (
0x90) Buffer-Overflow Injection (x86-64.text): AUC 0.9650 - Compiled Bytecode Payload Injection: AUC 1.0000 (native latent surprise AUC 0.9812)
- Single Bit-Flip Detection (RAM / Bytecode): AUC 0.9588 β 1.0000
Honest Scope & Findings
- KEEP consistently outperforms Transformers on raw binary sequences (2/2 wins) while using 2.4Γ fewer parameters and zero self-attention.
- On purely tabular datasets (EMBER), gradient-boosted trees remain the strongest baseline; on pure static byte-ngram classification, 1D CNNs remain highly competitive.
- KEEP TRACE's unique strength is combining zero-signature byte-level forensic localization (reducing SOC alert fatigue with exact offset ranges and benign-artifact triage) with sub-megabyte supervised ransomware classification (0.9988 AUC) in one linear-time $O(N \cdot k)$ architecture.
Β© 2026 Ameforge β Independent AI Research Studio, Windsor, Ontario.