Two undergraduate students from IIT Gandhinagar have developed an AI-powered security system designed to detect cyber attacks on water treatment plants, power grids and other industrial facilities before malicious digital commands turn into dangerous physical consequences. Called TrustGate, the prototype combines network activity with readings from physical sensors inside an industrial plant. Instead of merely alerting operators that something unusual is happening, the system is designed to identify the type of attack, locate the affected equipment and present actionable information in real time.

Third-year students Miloni Thakkar, from Computer Science and Engineering, and Mitwa Goswami, from Electrical Engineering, developed the project under the mentorship of Assistant Professor Sameer Kulkarni. Their work won second prize at the Intel Cup 2026, an international undergraduate electronic design competition held at Shanghai Jiao Tong University in China. The project demonstrates how artificial intelligence could potentially strengthen cyber security around systems that control essential services—but it is not yet ready for deployment at a real industrial facility.

IIT Gandhinagar TrustGate System Connects Cyber Activity With Physical Impact

Traditional industrial-security tools often focus heavily on network traffic. They can flag suspicious logins, malware, unusual communications or unauthorised attempts to access a system. The problem is that a command can sometimes look legitimate from a network perspective while still creating a dangerous physical change inside a plant. TrustGate attempts to close that gap. The system simultaneously studies two types of information: digital network traffic and readings produced by physical equipment such as sensors, pumps and valves.

This allows the model to ask a more important question than simply whether a computer command looks unusual: Does the digital command make sense given what is physically happening inside the facility?

That distinction could be particularly important in protecting critical infrastructure, where a malicious digital instruction can potentially affect water quality, electricity supply or industrial machinery.

The 2021 Florida Water Plant Attack Inspired the Project

The students partly based their idea on a widely reported 2021 incident at a water treatment facility in Oldsmar, Florida. According to the Press Information Bureau account, an unidentified attacker remotely accessed the plant's control system and attempted to increase the amount of sodium hydroxide used in water treatment to almost 100 times its normal safe level. The attempt was not detected automatically.

A plant operator noticed the computer cursor moving and reversed the change before contaminated water reached consumers. That example highlighted a problem for industrial cyber security systems. A command can be technically valid and therefore escape basic network-based detection, while its consequences in the physical world may be dangerous. TrustGate was designed around that exact problem.

How Does TrustGate Use Artificial Intelligence?

At the centre of the system is a dual-stream Bi-directional Long Short-Term Memory, or BiLSTM, model. One stream examines network behaviour. The second analyses data coming from physical sensors. A technique described by the team as cross-modal attention then allows the two streams to compare information with each other rather than making separate decisions.

In simple terms, imagine two investigators examining the same incident. One knows what is happening on the computer network. The other knows what pumps, valves and sensors are doing inside the plant. Instead of independently reporting their conclusions, the two continuously compare notes. That use of artificial intelligence may help the system detect slow or disguised attacks that do not look obviously malicious when network information or sensor readings are viewed separately.

TrustGate Can Point Operators Towards the Affected Equipment

Detecting that “something is wrong” is only part of the problem. During an attack, operators also need to know where the problem is happening and what action they may need to take.

TrustGate therefore attempts to identify the physical component most likely to have been affected. The information can then be displayed through a live map of the plant.

Thakkar explained in the official release that a generic warning may not be particularly useful during a real emergency. Operators need to know which component requires attention within seconds rather than spending valuable time searching through a complex facility. The team also carried out what it calls zero-shot testing, examining whether the system could identify attacks involving equipment it had not previously seen attacked during training.

Why the System Works Without the Cloud

Another unusual feature is that TrustGate has been designed to operate locally rather than depending on a cloud server. That matters because sensitive industrial networks are often deliberately isolated from the public internet. Connectivity may also be disrupted during a serious cyber incident. Relying on an external cloud service could therefore introduce both an additional attack surface and a potential point of failure.

The prototype instead runs entirely on embedded hardware. During the Intel competition, the students implemented the system on an Intel DK-2500 development platform. This local-processing approach could be particularly relevant to critical infrastructure, where operators often prioritise reliability and isolation over constant external connectivity.

Hardware Security Is Built Into the Prototype

The project does not rely solely on AI detection. TrustGate also incorporates TPM 2.0, or Trusted Platform Module technology. TPM can securely store cryptographic information and help verify that software and hardware have not been tampered with when a device starts. The prototype additionally includes a self-healing watchdog mechanism that can restart a critical process if it crashes or stops responding.

Goswami handled much of the hardware integration, TPM attestation and embedded Linux deployment, while Thakkar led development and testing of the AI architecture. The combination is intended to make the monitoring system itself more resilient rather than creating another vulnerable device inside an industrial network.

How Was the System Tested?

This is where an important limitation needs to be understood. TrustGate has not been tested inside an operating municipal water plant. The researchers used the Secure Water Treatment, or SWaT, dataset, developed from a scaled-down six-stage water-treatment testbed maintained by the iTrust Centre at the Singapore University of Technology and Design.

SWaT is widely used in industrial-control-system security research because it includes information from both normal plant operation and controlled cyber attack scenarios.

Using an established benchmark gives researchers a consistent environment in which different detection methods can be evaluated. But a laboratory dataset cannot capture every condition found inside a large operating facility.

What Must Happen Before TrustGate Can Be Used in the Real World?

The team and official release are clear that more work is required. Before the technology could be deployed, researchers would need to test it against live industrial data, add support for more control-system protocols and complete appropriate security certification. Real-world plants can contain decades-old equipment from multiple manufacturers.

Network configurations differ.

Sensor behaviour changes. False alarms are also a major concern because repeatedly warning operators about harmless anomalies can eventually cause real alerts to be ignored. These are problems that laboratory testing alone cannot resolve. The project should therefore be viewed as a promising research demonstration rather than a finished commercial defence system.

Why Industrial Cyber Security Is Different

Attacks on ordinary corporate IT systems primarily threaten information, accounts and business operations. Attacks on industrial systems can also create physical consequences. A compromised manufacturing controller can damage equipment. A malicious command in a power network can interrupt essential services. A manipulated water-treatment process could potentially threaten public health.

This makes industrial cyber security particularly demanding. Security software has to understand not only computers and networks, but also engineering processes and whether the physical behaviour of a plant makes sense. TrustGate's dual-data approach is an attempt to connect those two worlds.

Intel Cup Recognition Brings International Attention

The students had to demonstrate a working version of their system during the international Intel Cup competition. Participating teams were required to build on Intel's designated hardware, submit detailed technical documentation and defend their projects before an expert judging panel. Thakkar and Goswami secured second prize at the international competition in Shanghai.

The recognition gives the project greater visibility, but the more important next stage will be technical validation outside a competition environment. Moving from a successful academic prototype to a system trusted inside essential public infrastructure is a considerably larger engineering challenge. That means the project is not yet a finished shield against cyber attacks.

But it offers a practical glimpse of how the next generation of industrial-security systems may work: watching both the network and the physical world at the same time.