Skip to content

Energy & Utilities

Detecting changes in industrial control traffic

We tested similarity-preserving hashes for anomaly detection using a multi-vendor PLC bench and public labeled network captures.

Client
A federally funded research program, with a university SCADA testbed
Duration
12 months of research, through peer review
Team
4, with a national-laboratory review cycle
Year
2024–2025

Headline figures

Controller vendors validated against
4
Detection from deviation, not from a known-bad list
No signatures

The situation

What we walked into.

The detection tooling available for industrial environments largely inherits an assumption from enterprise security: that you know in advance what the bad thing looks like. In operational technology you frequently do not, because the attack that matters is often a legitimate protocol command issued at an illegitimate time by an unexpected party.

Endpoint agents were not an option. You cannot install software on a programmable logic controller, and the engineering workstation is frequently vendor-managed and out of scope for change.

Any active testing was off the table on the production process, which meant the research had to be conducted on a bench that faithfully represented several controller families rather than on a single convenient one.

There was also an honest scientific risk that the whole approach would not work. Similarity hashing is a technique from malware research, and the assumption that it would transfer usefully to process traffic was an assumption, not a result. It very nearly did not: industrial protocol frames are far smaller than the inputs these hash functions are designed for, and the first three approaches we tried failed outright.

Approach

How we sequenced it.

Each step had to be independently valuable. That constraint is what let the client stop at any point without being stranded.

  1. Build a bench that resembles a plant

    Controllers from four vendors, real ladder logic, and a process simulation that generated the traffic patterns a running line generates. Everything measured afterward depends on this being representative, so it got a disproportionate share of the effort.

  2. Characterize normal, in detail

    Passive capture over extended periods to establish what steady state actually looks like per device and per protocol, including the variation introduced by ordinary process changes, shift patterns and maintenance activity.

  3. Fingerprint with similarity hashes

    Similarity-preserving hashes over windows of clustered process traffic, so a small change produces a small distance rather than a completely different value. Clustering is what made it work at all: a single frame is too short for these functions to behave, and the cluster is the smallest unit that carries a stable signature. The output is a continuous deviation measure rather than a match or no-match verdict.

  4. Attack it deliberately, then measure

    Denial of service, man in the middle, packet injection and command manipulation, executed against the bench, plus evaluation against public capture sets carrying labeled attack windows. Results were presented for review at two national laboratories and written up for publication, which is a considerably harsher review than an internal report gets.

What was built

The parts that mattered.

  • Multi-vendor bench: Allen-Bradley, Siemens, Beckhoff and Click controllers with real ladder logic
  • Entirely passive detection: no agent on a controller, no scanning of the process network
  • Continuous deviation scoring rather than binary signature matching
  • Sliding window baseline that follows normal operational drift, so a maintenance window does not poison it
  • Attack classes validated by execution on the bench, not by literature review
  • Peer-reviewed at IEEE CARS 2024 and presented at two national laboratories

Results

What changed, and how we know.

3 of the 5 below are measurements, each against a stated baseline. The rest are states of the delivered system rather than numbers, and are written as such rather than dressed up as figures.

Controller families on the bench
4 vendors
Rather than a result demonstrated on one convenient platform
Installed on control equipment
0 agents
Detection is passive, off a span port
Network flows analyzed
100k+
Across bench capture and public labeled datasets
Method and results
Published
Peer-reviewed at IEEE CARS 2024, with the approaches that failed reported too
A primary signal, not a verdict
Indicator
It says something is off, not what, and belongs in front of a second-stage validator and a human
Built with
Python, Wireshark, TLSH, Modbus/TCP, Packet clustering, Elastic, Ladder logic

Next step

Tell us what’s breaking.

Forty-five minutes, no charge. We’ll say what we’d do, what it would likely cost, and whether the thing you’re worried about is actually your biggest risk.

Reply
A person replies, not a sequence: within one business day, from someone who would be on the engagement.