Only learn from data the host can vouch for.
ExVCL is a Python library that sits in front of a continual learner. It checks each labelled observation against execution evidence recorded on the host, then admits, rejects or defers it, so poisoned labels never reach the model update.
# install and verify $ pip install exvcl $ python -m exvcl selftest # 37/37 tests $ python -m exvcl demo # end-to-end check on synthetic evidence
// problem
The training stream is an attack surface
Continual learners keep updating from new data. If an attacker controls even some of the labels, a backdoor can be implanted.
Labels can be attacked
A few flipped labels reaching the update path can change what the model learns.
Data-only filters miss behaviour
Sanitisers inspect the data and its labels. They cannot tell whether the session behaved the way its label claims.
ExVCL checks the behaviour
Evidence recorded from the host execution is compared with the label before the model may learn from it.
// architecture
How it fits together
ExVCL is an admission layer. Your upstream model, continual learner, replay memory and objective stay unchanged.
// modes
Three admission modes
| Mode | Behaviour | Main consideration |
|---|---|---|
none | Accepts everything; no verification. | Accept-all baseline for measuring harm. |
selective | Verifies observations picked by model and label triage; may audit part of the fast path. | Needs attack scores and triage calibration; unverified fast-path rows can be admitted. |
strict | Verifies observations within the verification budget; the rest are deferred, not learned. | No fast path; a low budget leaves less data to learn from. |
// live demo
Replay a real run
400 sessions recorded on two Google Cloud VMs, with 20% of labels flipped by a simulated attacker. Pick a gate and watch what reaches the learner.
Session detail (click any square)
b40-s9
ssh_bruteforce (malicious)malicious
upstream model said maliciousmalicious
P(malicious) = 0.99ADMIT
evidence agrees with the labelDoes the evidence matter? Same stream replayed, 20 gate seeds
| Gate | Poison admitted (of 76) | Correct labels kept | Macro-F1 |
|---|---|---|---|
| Accept-all | 76.0 | 100% | 0.940 |
| Selective, real evidence, 100% budget | 6.0 | 88.5% | 0.947 |
| Strict, real evidence, 30% budget | 1.8 | 26.3% | 0.947 |
| Strict, real evidence, 50% budget | 3.0 | 44.2% | 0.947 |
| Strict, real evidence, 70% budget | 4.4 | 61.8% | 0.947 |
| Strict, real evidence, 100% budget | 6.0 | 88.2% | 0.947 |
| Strict, shuffled evidence (control) | 37.9 | 49.1% | 0.942 |
| Strict, global-random evidence (control) | 37.0 | 50.1% | 0.934 |
With the evidence scrambled, the gate admits about six times more poison and throws away about half of the good data. The protection comes from what the host actually did.
Known blind spot. All 6 poisoned labels that Strict admitted are port scans flipped to benign: host evidence reads a port scan as benign. This was a controlled lab run on seven session types, so it does not show robustness on arbitrary real-world traffic.
// quick start
Use it in your pipeline
Fit the evidence model on a trusted labelled calibration set, bootstrap the learner, then pass each chunk through the gate.
import numpy as np
from exvcl import EvidenceModel, ExVCL
ev = EvidenceModel(target_fpr=0.02, target_fnr=0.02).fit(F_cal, y_cal_malicious)
model = ExVCL(n_features=X0.shape[1], mode="strict", evidence=ev,
learner="softmax", budget=1.0)
model.bootstrap(X0, y0) # trusted initial data
verify = lambda idx: (F_chunk[idx], np.ones(len(idx), dtype=bool))
result = model.process_chunk(X_chunk, y_observed, verify=verify)
print(result.summary())The seven evidence counts, in fixed order:
process_spawnshell_execfile_writesensitive_file_accessnet_connect_outpriv_changesyscall_total// scope
What to know before you use it
Research framework
Not a turnkey host-security product. Results depend on telemetry source, coverage, calibration, learner and budget. Evaluate it in your environment first.
Known blind spot
Port scans read as benign to host evidence in our setup. In the demo, every admitted poisoned label is a flipped port scan.
Controlled lab
The demo run used scripted sessions of seven types. It is not a claim about arbitrary real-world traffic.
// related
Related
Try it on your own telemetry.
Install it, run the self-test, then point the gate at your own evidence.