NonSense: On the (In)security of FPGA Anti-tamper Sensors

arXiv:2611.17065 · cs.CR · Submitted 2026-11-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "NonSense: On the (In)security of FPGA Anti-tamper Sensors".

Elias: Here's the gist of the paper. FPGAs are widely used in mission-critical systems where resilience to physical attacks is a critical security requirement To defend against active physical attacks, such as glitching attacks, modern FPGAs rely on on-chip sensors that detect these disturbances and trigger responses, such as key zeroization, before a device is compromised This work challenges the assumption that the sensors themselves are a fragile and underexplored attack surface by showing both detection and response mechanisms can be bypassed without invasive tampering The authors argue that sensor reliability is reduced by factors rarely captured in threat models, including spatial and frequency-dependent behavior of the system’s power delivery network, environmental conditions such as temperature, and long-term aging They propose countermeasures like asynchronous sensing and response paths to harden detection and response mechanisms against these effects.

Nadia: Sensor Vulnerabilities Under Environmental Stress The reliability of sensors is affected by factors beyond typical threat models, such as the spatial and frequency-dependent behavior of the system’s power delivery network, which can steer a fault to a target while sparing nearby sensors Environmental conditions like temperature shift detection thresholds, and long-term aging drifts sensor calibration and may open silent detection gaps The authors characterize sensor behavior under accelerated aging, measuring changes in RO and TDC outputs and the loss of valid XADC temperature readings Specifically, for soft-IP sensors subjected to accelerated aging, the RO frequency decreases from approximately 202.3 MHz on a fresh device to 197 MHz after 535 hours, a total reduction of ∼2.6% Similarly, the TDC delay tap count falls from 58.0 taps on the fresh device to 32.4 taps after 535 hours, a 44% reduction.

Elias: Characterization of Hard IP Sensor Failures The study examines failures in hard-IP sensors under extreme conditions, revealing that plausible sensor readings do not establish continued measurement availability when cooling occurs For the XADC temperature output during accelerated aging, the reported temperature increased from 30.5°C on a fresh device to 32.2°C after 3 hours, and then returned code 0x000 corresponding to −273.15°C after 36 hours Furthermore, under liquid-nitrogen cooling, the TVS returned invalid readings while clock activity remained observable during part of the cooling process.

Nadia: These results show that hard-IP monitoring and its supporting infrastructure can become unreliable under cooling while sensitive state remains retained.

Paper summary: Nadia: So, we're looking at this paper called "NonSense: On the (In)security of FPGA Anti-tamper Sensors," which tackles how these on-chip sensors designed to stop physical attacks, like glitching attacks, can actually be vulnerable themselves. The authors argue that we often take these sensors for granted as being reliable, but they show that both the detection and the response parts of these mechanisms can be bypassed without needing a major physical tamper.

Elias: It's interesting because they challenge that assumption directly; they don't assume the sensors are perfect or immune to attack themselves, which is a big shift in how we think about FPGA security, especially since FPGAs are so common in mission-critical systems.

Priya: So if the sensor can be bypassed, what does that mean for the actual security goal of zeroizing sensitive data? Does it just mean the zeroization doesn't happen when it should?

Nadia: Exactly; they show that we need to look at how environmental stress and physical layout mess with those detection and response stages, not just a simple attack vector on top of them. The paper focuses on how things like temperature and aging affect the outputs of these sensors, which is something we rarely model in our threat models.

Elias: They point out that the reliability of these sensors is affected by things like the spatial and frequency-dependent behavior of the power delivery network, which can steer a fault to one sensor while leaving others untouched. That's a key detail about how physical placement matters.

Priya: What kind of effects are they seeing when they look at that spatial dependency? Is it just that some sensors get hit harder because of their location?

Nadia: It goes deeper than just location; the temperature variations themselves cause spatial dependence, meaning some parts of the FPGA get faster or slower depending on where you are located. It’s not uniform across the device.

Elias: They measured concrete numbers for aging effects too; they showed that for soft-IP sensors subjected to accelerated aging, the ring oscillator frequency drops from about two hundred two point three megahertz on a fresh chip to about one hundred ninety-seven megahertz after five hundred thirty-five hours, which is a reduction of about two point six percent.

Paper summary: Priya: That drop in the clock speed sounds like it could cause problems for timing-based detection circuits, so how does that tie into the overall system failure?

Nadia: It’s tied to how it affects the time-to-digital converter, or TDC, which they found fell from fifty eight point zero taps on a new device down to thirty two point four taps after five hundred thirty-five hours, which is a forty four percent reduction.

Elias: And they also looked at failures in hard-IP sensors under extreme conditions; they found that plausible sensor readings don't actually mean the measurement keeps going once cooling occurs. For instance, the temperature output from the XADC during accelerated aging went up to three point two degrees Celsius after three hours, and then it returned a code zero zero zero corresponding to minus two hundred seventy-three point one five degrees Celsius after thirty six hours.

Priya: So they are saying that even if the sensor *reports* a reading, that reading might be completely invalid or misleading when the environment changes? That sounds like a major trust issue.

Nadia: It is; and under liquid nitrogen cooling, they saw the TVS returned invalid readings while the clock activity stayed observable for part of that cooling process. This suggests that the hard-IP monitoring and all its supporting infrastructure can become unreliable when things get cold, even if the sensitive state itself stays retained.

Elias: They also characterized how placement and routing affect delay-based sensors, like TDCs and ring oscillators, showing that even at a fixed temperature, the behavior isn't uniform across the FPGA; some sensors run faster while others run slower. This reinforces that proximity doesn't guarantee equal exposure to a disturbance.

Priya: So if we are designing these systems, what kind of practical countermeasures are they suggesting to deal with this combination of aging and spatial variation?

Nadia: They propose things like asynchronous sensing and response paths, plus sensor health monitoring, bounded environmental compensation, and controlled recalibration under trusted conditions. They argue that detection thresholds need margins for temperature variation and aging with some form of bounded recalibration when the system is still trusted.

Elias: They also stress that independent health checks should look for invalid readings or stalled conversions instead of just relying on the sensor's own alarm, and placement and routing needs to be characterized alongside the protected logic because proximity alone isn't enough to guarantee equal exposure.

Paper summary: Priya: That moves us from just fixing the sensor to redesigning the whole sensing-to-clearing path based on these environmental factors; what does that mean for system design?

Nadia: It means robust protection requires you to characterize the complete sensing-to-clearing path under combined attacks and environmental variation, because a sensor might detect an attack but fail to protect if its response needs a clock edge that the attack has already suppressed.

Elias: They conclude that asynchronous sensing and response can strengthen protection against those clock disruption attacks because it lets the alarm generation and sensitive state clearing happen without needing another edge of the monitored clock. However, they also make a key point about limitations: an autonomous sensing element doesn't automatically make the whole sensor asynchronous if parts like measurement or threshold comparison still depend on a vulnerable clock.

Priya: So if we look at what this paper actually shows us, it seems like the real security isn't just in having a sensor, but in making sure that when an attack happens, the entire sequence—detecting it and wiping the state—can happen reliably throughout its intended operating lifetime.

Nadia: That’s right; they are forcing us to assess anti-tamper mechanisms by whether they actually remove sensitive state over the whole intended operating lifetime. This work, "NonSense: On the (In)security of FPGA Anti-tamper Sensors," tells us that we need to look beyond just the immediate attack vector and account for the physical reality of how these chips age and change under temperature stress.

Elias: It’s a necessary caution because it shows that relying on a sensor's simple alarm isn't enough if the underlying physics of how that sensor works is being subtly altered by its own environment or its placement.

Priya: So for anyone building these systems, the implication is to be very specific about what kind of environmental stress you can tolerate and build in extra robustness against those known physical dependencies, like spatial variations in delay.

Nadia: That’s the practical side of it; we have to move toward designs that include things like bounded compensation and independent health checks because just having a sensor is not security anymore if its reliability changes with the temperature or how you wired it.

Conclusion: Nadia: So we've been looking at how these on-chip sensors meant to stop physical attacks can actually be fooled by things like aging and temperature changes, and now we're wrapping up this paper called "NonSense: On the (In)security of FPGA Anti-tamper Sensors."

Elias: Yeah, it’s about challenging the idea that these sensors are perfectly reliable. The authors are showing how even if you have a sensor that detects a glitch, the way it operates under stress—like temperature shifts or long-term aging—can break its detection and response cycle.

Priya: So what is the core message here? Like, if we rely on these sensors to keep data safe in critical systems, are they actually safe over time?

Nadia: The main point is that sensor reliability isn't just about the immediate attack; it’s about how much physical stress—like the way electricity flows through the chip or how hot it gets—reduces that reliability. It’s not a simple security failure; it’s a physics problem.

Elias: They propose things like asynchronous sensing and response paths because they think those can help harden those detection mechanisms against clock disruptions. But they also point out that if the part of the sensor that does the math still relies on a vulnerable clock, it doesn't fix everything.

Priya: So for someone who just listens to this show, what does this mean for a system designer? Does it mean we have to add extra hardware just to compensate for chip aging?

Nadia: It means you can’t just trust the alarm from one sensor alone. You need independent health checks that look at invalid readings or stalled conversions instead of relying on the sensor's own alarm. And you have to design in margins for temperature variation and aging right into your timing thresholds.

Elias: Exactly, it forces a redesign of how we think about protection. It’s not enough to detect an attack; you have to make sure the whole path from sensing that disturbance to wiping the sensitive state works reliably throughout the entire life of the device.

Priya: So for now, it seems like we’re shifting from asking "can this chip stop this attack?" to "can this chip keep working reliably while stopping this attack over years?"

Nadia: Right. And that leads us right into how we actually test these kinds of long-term reliability issues in the real world.

Mehmet Ali Cetin, Fatemeh Khojasteh Dana, Saleh Khalaj Monfared, Nima Seyed Aghazadeh, Maryam Saadat Safa, Shahin Tajik

Worcester Polytechnic Institute

cs.CR

Submitted: 2026-11-12

Updated: 2026-11-12

Code: https://github.com/0xADE1A1DE/Borrowed-Time

Importance score: 81/100

The gist: FPGAs are widely used in mission-critical systems where resilience to physical attacks is a critical security requirement To defend against active physical attacks, such as glitching attacks…

Key concepts

Environmental Stress Effects
Sensors are unreliable due to real-world conditions beyond simple threat models. Temperature shifts detection thresholds, and aging causes calibration drifts, leading to gaps where attacks go undetected.
Spatial and Frequency Dependence
The physical placement and routing of sensors matter significantly. Some sensors behave differently based on their location on the chip or the frequency of signals they monitor. This means a disturbance might only affect certain sensors depending on where they are located.
Asynchronous Sensing and Response
This is a proposed defense mechanism where detection and response paths operate independently of the main system clock. This prevents attacks that specifically target clock edges from stopping the alarm or clearing sensitive data.
Complete Protection Path
Security depends on the entire sequence from detecting an attack to successfully wiping sensitive state during the device's entire lifespan. A sensor alone is insufficient; its response mechanism must also be robust against all variations.

Terminology

Summary

FPGAs are widely used in mission-critical systems where resilience to physical attacks is a critical security requirement To defend against active physical attacks, such as glitching attacks, modern FPGAs rely on on-chip sensors that detect these disturbances and trigger responses, such as key zeroization, before a device is compromised This work challenges the assumption that the sensors themselves are a fragile and underexplored attack surface by showing both detection and response mechanisms can be bypassed without invasive tampering The authors argue that sensor reliability is reduced by factors rarely captured in threat models, including spatial and frequency-dependent behavior of the system’s power delivery network, environmental conditions such as temperature, and long-term aging They propose countermeasures like asynchronous sensing and response paths to harden detection and response mechanisms against these effects. p. 0

Sensor Vulnerabilities Under Environmental Stress

The reliability of sensors is affected by factors beyond typical threat models, such as the spatial and frequency-dependent behavior of the system’s power delivery network, which can steer a fault to a target while sparing nearby sensors Environmental conditions like temperature shift detection thresholds, and long-term aging drifts sensor calibration and may open silent detection gaps The authors characterize sensor behavior under accelerated aging, measuring changes in RO and TDC outputs and the loss of valid XADC temperature readings Specifically, for soft-IP sensors subjected to accelerated aging, the RO frequency decreases from approximately 202.3 MHz on a fresh device to 197 MHz after 535 hours, a total reduction of ∼2. p. 36% Similarly, the TDC delay tap count falls from 58. p. 30 taps on the fresh device to 32. p. 34 taps after 535 hours, a 44% reduction. p. 3

Characterization of Hard IP Sensor Failures

The study examines failures in hard-IP sensors under extreme conditions, revealing that plausible sensor readings do not establish continued measurement availability when cooling occurs For the XADC temperature output during accelerated aging, the reported temperature increased from 30.5°C on a fresh device to 32. p. 32°C after 3 hours, and then returned code 0x000 corresponding to −273. p. 315°C after 36 hours Furthermore, under liquid-nitrogen cooling, the TVS returned invalid readings while clock activity remained observable during part of the cooling process. p. 4 These results show that hard-IP monitoring and its supporting infrastructure can become unreliable under cooling while sensitive state remains retained. p. 5

Spatial and Temporal Dependencies

The analysis demonstrates that placement and routing significantly affect FPGA-based sensors, particularly delay-based sensors such as TDCs and ROs Even at a fixed temperature, the behavior of the TDCs is not spatially uniform across the FPGA, with some sensors operating faster while others are slower The spatial and frequency-dependent behavior reported in GlitchSnipe further indicates that a nearby sensor need not experience the same supply disturbance as the protected logic Temperature variations also exhibit spatial dependence, with some locations becoming faster and others slower across the device.

Proposed Design Recommendations

The work derives design recommendations for more resilient protection, including asynchronous sensing and response paths, sensor-health monitoring, bounded environmental compensation, and controlled recalibration First, detection thresholds and response timing require margins for temperature variation and aging with bounded recalibration under trusted conditions Second, independent health checks should identify invalid readings and stalled conversions rather than rely exclusively on a sensor’s own alarm Third, placement and routing should be characterized jointly with the protected logic because proximity alone does not ensure equal exposure to a disturbance. Robust protection requires characterization of the complete sensing-to-clearing path under combined attacks and environmental variation. p. 6

Conclusion on Protection Strategy

The security of FPGA anti-tamper mechanisms depends on the complete path from sensing a disturbance to wiping the sensitive state A sensor may correctly detect an attack yet provide no protection if its response requires clock edges that the attack suppresses Asynchronous sensing and response can strengthen protection against clock-disruption attacks by allowing alarm generation and sensitive-state clearing to proceed without further edges of the monitored clock However, an autonomous sensing element does not make the complete sensor asynchronus if measurement, threshold comparison, or alarm propagation still depends on a vulnerable clock A clock-stop detector also does not necessarily detect a short glitch that corrupts computation while the clock continues to run The paper concludes that FPGA anti-tamper protection should ultimately be assessed by whether it removes sensitive state in time throughout its intended operating lifetime.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can incorporate asynchronous sensing and response paths to strengthen protection against clock-disruption attacks by allowing alarm generation and sensitive-state clearing to proceed without further edges of the monitored clock.

  2. The improved system can implement a delaybased clock-stop detector that asserts an alarm without waiting for another edge of the monitored clock, while asynchronous reset or preset paths can overwrite appropriately implemented sensitive registers after that clock stops.

  3. Robust protection requires characterization of the complete sensing-to-clearing path under combined attacks and environmental variation, ensuring security is assessed by whether it removes sensitive state in time throughout its intended operating lifetime.

  4. The system should implement independent health checks to identify invalid readings and stalled conversions rather than relying exclusively on a sensor's own alarm, addressing the finding that a low reported code need not trigger a high-temperature alarm.

  5. Design recommendations should include bounded recalibration under trusted conditions to account for substantial changes in sensor baselines observed during accelerated aging.

Sources

Related papers