DeepChronos: Temporal Neuro-Symbolic Reasoning for Multi-Step Attack Detection
Organizations: Delft University of Technology · University of Turku
Abstract
Modern cyber threats commonly unfold as multi-step attacks comprising temporally and causally related phases. Current defenses commonly follow a detect first and correlate later paradigm. Intrusion detection systems classify events independently, while downstream systems reconstruct attack progression from the resulting alerts or audit records. In this context, learning-based detectors learn attack-step dependencies implicitly as statistical regularities in the data and cannot guarantee that their predictions follow the real attack progression. Conversely, correlation systems represent attack relationships explicitly, but apply this knowledge only after the underlying attack observations have been detected. We present DEEPCHRONOS, a neuro-symbolic network intrusion detection system that shifts the paradigm from detect first and correlate later to detection through logical progression. DEEPCHRONOS combines learned flow-level evidence with explicit attack-phase rules, temporal prerequisites, and previously inferred phases, allowing the evolving attack history to inform and constrain each detection decision. Evaluations on DARPA 2000 and AIT-LDS V2 show that DEEPCHRONOS improves macro F1 over purely neural baselines by up to 11.3%, eliminates causal violations under strict temporal constraints, and maintains millisecond-scale inference latency. These results demonstrate that integrating explicit attack progression knowledge into detection produces more accurate and temporally coherent multi-step attack predictions.
Figures & tables
| Model | Macro F1 ( std) | Weighted F1 | Accuracy | DR | FAR | CVR | RVR |
|---|---|---|---|---|---|---|---|
| AIT-LDS V2 | |||||||
| DeepChronos Pretrained T+R | 0.944 0.024 | 0.999 | 0.999 | 0.990 | 0.0003 | 0.000 | 0.002 |
| DeepChronos E2E T+R | 0.940 0.010 | 0.999 | 0.999 | 0.991 | 0.0005 | 0.000 | 0.014 |
| LSTM Multiclass | 0.880 0.015 | 0.998 | 0.998 | 0.988 | 0.0018 | 0.019 | 0.021 |
| DeepChronos Pretrained T | 0.878 0.020 | 0.998 | 0.998 | 0.983 | 0.0008 | 0.000 | 0.152 |
| LSTM Ensemble | 0.858 0.024 | 0.998 | 0.998 | 0.991 | 0.0019 | 0.027 | 0.017 |
| Model | Macro F1 | Weighted F1 | F1 Benign | F1 P1 | F1 P2 | F1 P3 | F1 P4 |
|---|---|---|---|---|---|---|---|
| DeepChronos E2E | 0.721 | 0.945 | 0.997 | 0.977 | 0.752 | 0.877 | 0.026 |
| DeepChronos Pretrained | 0.639 | 0.921 | 0.962 | 0.981 | 0.620 | 0.585 | 0.046 |
| LSTM Multiclass | 0.596 | 0.898 | 0.997 | 0.965 | 0.534 | 0.511 | 0.024 |
| Logic Baseline | 0.561 | 0.953 | 0.982 | 0.000 | 0.999 | 0.824 | 0.000 |
| Model | Macro F1 | Weighted F1 | F1 Benign | F1 P1 | F1 P2 | F1 P3 | F1 P4 | F1 P5 |
|---|---|---|---|---|---|---|---|---|
| Logic Baseline | 0.931 | 0.871 | 0.686 | 0.998 | 1.000 | 1.000 | 0.974 | 0.926 |
| DeepChronos E2E | 0.912 | 0.974 | 0.951 | 0.586 | 1.000 | 1.000 | 0.947 | 0.990 |
| DeepChronos Pretrained | 0.886 | 0.971 | 0.947 | 0.431 | 1.000 | 0.974 | 0.973 | 0.991 |
| LSTM Multiclass | 0.410 | 0.104 | 0.380 | 0.468 | 0.367 | 0.904 | 0.333 | 0.007 |
| Configuration | Macro F1 | Weighted F1 | F1 Benign | F1 P1 | F1 P2 | F1 P3 | F1 P4 |
|---|---|---|---|---|---|---|---|
| Logic Orig , Data Orig | 0.931 | 0.999 | 0.999 | 0.988 | 1.000 | 0.769 | 0.900 |
| Logic Orig , Data Corr | 0.899 | 0.994 | 0.997 | 0.989 | 0.822 | 0.769 | 0.920 |
| Logic Rev , Data Corr | 0.917 | 0.998 | 0.999 | 0.987 | 0.986 | 0.857 | 0.754 |
| Model | Latency (ms) | Throughput (pred/s) | Peak memory (GB) |
|---|---|---|---|
| DeepChronos Pretrained T | 403 | ||
| DeepChronos E2E T | 415 | ||
| DeepChronos Pretrained T+R | 2155 | ||
| DeepChronos E2E T+R | 2152 | ||
| Logic Baseline T+R | 4354 | ||
| Logic Baseline T | 6985 |
| Model | Latency (ms) | Throughput (pred/s) | Peak memory (GB) |
|---|---|---|---|
| DeepChronos Pretrained T | 189 | ||
| DeepChronos E2E T | 187 | ||
| DeepChronos Pretrained T+R | 178 | ||
| DeepChronos E2E T+R | 189 | ||
| Logic Baseline T+R | 327 | ||
| Logic Baseline T | 457 |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Scenario | Simulation period | Attack period | Scan volume | Size |
|---|---|---|---|---|
| fox | 2022-01-15 to 2022-01-20 | 2022-01-18 11:59–13:15 | High | 26 GB |
| santos | 2022-01-14 to 2022-01-18 | 2022-01-17 11:15–11:59 | Low | 17 GB |
| Sensor | Benign | Phase 1 | Phase 2 | Phase 3 | Phase 4 | Phase 5 |
|---|---|---|---|---|---|---|
| Inside | 52,258 | 20 | 22 | 34 | 19 | 73,472 |
| DMZ | 10,609 | 767 | 25 | 79 | 19 | 33,942 |
| Scenario | Benign | Phase 1 | Phase 2 | Phase 3 | Phase 4 |
|---|---|---|---|---|---|
| fox | 400,024 | 14,678 | 56,443 | 28 | 61 |
| santos | 333,964 | 12,861 | 5,670 | 26 | 228 |
| Attribute | Description | Usage |
|---|---|---|
| ts | Connection timestamp | Symbolic |
| id.orig_h | Source address | Symbolic |
| id.resp_h | Destination address | Symbolic |
| id.orig_p | Source port | Symbolic |
| id.resp_p | Destination port | Symbolic |
| local_orig | Internal source indicator | Symbolic + Neural |
| Type | Features |
|---|---|
| Categorical | local_orig , local_resp , proto , service , conn_state , and ip_proto . |
| Numerical | duration , orig_bytes , resp_bytes , missed_bytes , orig_pkts , orig_ip_bytes , resp_pkts , resp_ip_bytes , connections_per_src , unique_targets , unique_dports , connections_per_dst , unique_sources_per_dst , syn_failure_ratio , reject_ratio , and reset_ratio . |
| Configuration | Evaluated options |
|---|---|
| Feature sets | Base features, flow-only features |
| Window sizes | 10, 100 |
| Learning rates | , |
| Model | Features | Window | Learning rate | Macro F1 | Accuracy | DR | FAR |
|---|---|---|---|---|---|---|---|
| LSTM Ensemble | Base | 100 | 0.8650 | 0.9966 | 0.9874 | 0.0029 | |
| LSTM Multiclass | Base | 100 | 0.8213 | 0.9952 | 0.9865 | 0.0043 | |
| LSTM Ensemble | Base | 100 | 0.7498 | 0.9907 | 0.9867 | 0.0090 | |
| LSTM Multiclass | Base | 100 | 0.6919 | 0.9874 | 0.9846 | 0.0124 | |
| LSTM Multiclass | Base | 10 | 0.6749 | 0.9923 | 0.9756 | 0.0067 | |
| LSTM Ensemble | Base | 10 | 0.6741 | 0.9924 | 0.9760 | 0.0066 |
| Model | Features | Window | Learning rate | Macro F1 | Accuracy | DR | FAR |
|---|---|---|---|---|---|---|---|
| LSTM Ensemble | Base | 10 | 0.9152 | 0.99958 | 0.99957 | 0.00030 | |
| LSTM Multiclass | Base | 10 | 0.9142 | 0.99957 | 0.99953 | 0.00027 | |
| LSTM Ensemble | Flow-only | 10 | 0.9131 | 0.99955 | 0.99959 | 0.00041 | |
| LSTM Multiclass | Flow-only | 10 | 0.8850 | 0.99949 | 0.99955 | 0.00046 | |
| LSTM Ensemble | Base | 100 | 0.8843 | 0.99955 | 0.99953 | 0.00033 | |
| LSTM Ensemble | Flow-only | 100 | 0.8817 | 0.99953 | 0.99951 | 0.00030 |
Network Intrusion Detection
Network intrusion detection identifies malicious or abnormal network activity to help protect computers, connected devices, and critical infrastructure.
See all 181 papers