
Detecting infected machines on a corporate network is a problem the industry has been trying to solve for decades. Each generation of tools attacked it from a different angle, and each one added its own stream of alerts. The result, in many security teams, is not a lack of data but a flood of it, most of it false.
Layers that grew one at a time
First came the antivirus agent on each machine, scanning files for known signatures. It worked well, and still does for known malware. When computers became reachable from anywhere, firewalls appeared to filter incoming connections, and network address translation separated private networks from the internet. Intrusion detection and prevention systems followed, inspecting traffic for attack patterns. Later still came sandboxes, endpoint detection and response, and security information and event management (SIEM) platforms to collect it all.
Every layer was a reasonable answer to a real problem. Together, they often produce thousands of alerts a day, and analysts learn to ignore most of them. That is how genuine intrusions get missed: the alert existed, but it looked like all the others.
What "orthogonal" means here
Two detection methods are orthogonal when they look at independent evidence. An antivirus signature and a network IDS rule written from the same malware sample are not orthogonal; they will tend to fire, or stay silent, together. A file signature, an unusual login pattern and a touch on a decoy server are orthogonal, because each one observes a different aspect of the attack.
When orthogonal detectors agree, confidence rises sharply. A single anomaly score might be wrong one time in ten. Two independent signals pointing at the same host at the same time are far less likely to be a coincidence. This is the basic idea behind correlation rules in a SIEM, and behind risk-based alerting, where small signals add up until a host crosses a threshold.
Deception as a near-zero false positive layer
Most detection methods have to separate good behaviour from bad. Deception avoids that problem. A decoy server, a fake credential or a planted document has no business purpose, so legitimate users have no reason to touch it. When something does, the alert is usually meaningful on its own, and it becomes very strong when combined with another signal from the same host.
| Signal | What it observes | Typical false positive source |
|---|---|---|
| Antivirus signature | File content | Generic or overly broad signatures |
| Behavioural analytics | User and host activity over time | New projects, travel, role changes |
| Network detection | Traffic patterns and protocols | Unusual but legitimate software |
| Deception (decoys, honeytokens) | Interaction with planted assets | Mostly scanners and admin tools you forgot to exclude |
Putting it into practice
Start by listing your detection sources and asking, for each pair, whether they really observe different things. Then write correlation rules that require at least two independent sources before paging a human at night. Keep the single-source alerts, but route them to a queue that is reviewed in daylight. Finally, add a small deception layer, because it fills the gap that most of the other tools share: they are weakest exactly when an attacker is quietly moving around with stolen credentials.
The same thinking is used well outside corporate IT. Fraud teams at online sportsbooks, for example, rarely act on one suspicious bet; they look for independent signals, such as device, payment method and betting pattern, that agree before they freeze an account. Fewer false alarms means the real ones get attention.