README.md

Secure Hardware Design Guidelines

Overview

Silicon designs for security devices require special guidelines to protect the designs against myriad attacks. For OpenTitan, the universe of potential attacks is described in our threat model. In order to have the most robust defensive posture, a general approach to secure hardware design should rely on the concepts of (1) defense in depth, (2) consideration of recovery methods post-breach, and (3) thinking with an attacker mindset.

In all cases, as designers, we need to think of equalizing the difficulty of any particular attack for the adversary. If a design has a distribution of attack vectors (sometimes called the "attack surface" or "attack surface area"), it is not the strength of the strongest defenses that is particularly of interest but rather the weakest, since these will be the most likely to be exploited by the adversary. For example, it's unlikely that an attacker will try to brute-force a system based on AES-128 encryption, as the difficulty level of such an attack is high, and our confidence in the estimate of the difficulty is also high. But, if the security of the AES-128 depends on a global secret, more mundane attacks like theft or bribery become more likely avenues for the adversary to exploit.

Defense in depth means having multiple layers of defenses/controls acting independently. Classically in information security, these are grouped into three main categories: physical, technical and administrative¹. We map these into slightly different elements when considering secure hardware design:

• Physical security typically maps to sensors and shields, but also separation of critical information into different locations on the die.
• Technical security includes techniques like encrypting data-at-rest, scrambling buses for data-in-motion, and integrity checking for all kinds of data.
• Administrative security encompasses architectural elements like permissions, lifecycle states, and key splits (potentially also linked to physical security).

Consideration of recovery methods means assuming that some or all of the defenses will fail, with an eye to limiting the extent of the resulting system failure/compromise. If an adversary gains control over a sub-block, but cannot use this to escalate to full-chip control, we have succeeded. If control over a sub-block is detected, but an alert is generated that ultimately causes a device reset or other de-escalation sequence, we have created a recovery strategy. If the software is compromised but access to keys/secrets is prevented by hardware controls, we have succeeded. If compromise of secrets from a single device cannot be leveraged into attacks on other (or all) devices, again we have succeeded. If compromised devices can be identified and quarantined when enrolled into a larger system, then we have a successful recovery strategy.

Thinking with an attacker mindset means "breaking the rules" or violating assumptions: what if two linked state machines no longer are "in sync" - how will they operate, and how can they recover? What happens if the adversary manipulates an internal value (fault injection)? What happens if the adversary can learn some or all of a secret value (side channel leakage)? This document will primarily try to give generic guidance for defense against the latter two attacks (fault injection, and side channel information leakage). It also discusses ways to either prevent attacks, mitigate them, or alert of their existence. Other attack vectors (especially software compromises or operational security failures) are not in the scope of this document, or will be addressed at a later stage.

In general, when thinking of protecting against fault injection attacks, the designer should consider the consequences of any particular net/node being inverted or forced by an adversary. State of the art fault attacks can stimulate two nodes in close succession; robustness to this type of attack depends on the declared threat model. Designers need to be well aware of the power of an attack like SIFA [15], which can bypass "conventional" fault countermeasures (e.g. redundancy/detectors) and requires only modest numbers of traces.

For increased resistance against side channel leakage (typically: power, electromagnetic radiation, or timing), designs in general should ensure that the creation or transmission of secret material is handled in such a way as to not work with "small" subsets of bits of sensitive information. All side channel attacks use a strategy of "divide and conquer" to recover secret material e.g. word by word or byte by byte. Making sure to evaluate/process information in 32-bit quanta or larger will make this more difficult. Attacks like DPA are especially powerful because on top of such a "divide and conquer" strategy, where e.g. in the case of AES DPA addresses algorithm-inherent 8-bit intermediate values, they benefit from statistical analysis of a large number of side channel observations to recover the correct values.

Below we will go deeper into these recommendations for general design practices. Individual module guidance for particular IP (processor, AES, SHA, etc) will be handled in addenda to this document.

General Module Level Design Guidance

These guidelines are for sub-block / module level design. System architecture, identity management, and protocol design are outside the scope of this document, but may create some dependencies here. For general reading, the slides of [10] are considered a useful companion to these guidelines.

Recommendation 1: Identify sensitive/privileged operations

Identify any sensitive/privileged operations performed by the module (non-exhaustive list of examples: working with secret keys, writing to OTP, potentially writing to flash, enabling debug functionality, lifting/releasing access restrictions, changing lifecycle state)

1. Having these operations documented helps to analyze the potential issues of any attack discussed below.
2. Subsequent design/verification reviews can use these sensitive operations as focus areas or even coverage points.

Recommendation 2: Side-channel leakage considerations

Consider side-channel leakage of any secret information (side channels include timing, power, EM radiation, caches, and micro-architectural state, among others)

1. Process secret information in at least a 32-bit wide datapath
2. Use fixed/constant time operations when handling secrets (see [6] and [11])
3. Don't branch/perform conditional operations based on secret values
4. Incorporate temporal randomness (example: add delay cycles based on LFSR around critical operations, see [9])
5. Cryptographic operations should incorporate entropy (via masking/blinding, see [9]), especially if the key is long-lived, or a global/class-wide value. Short-lived keys may not require this, but careful study of the information leakage rate is necessary
6. Noise generation - run other "chaff" switching actions in parallel with sensitive calculations, if power budget permits (see [9])
7. Secrets should not be stored in a processor cache (see [3])
8. Speculative execution in a processor can lead to leakage of secrets via micro-architectural state (see [4]/[5])
9. When clearing secrets, use an LFSR to wipe values to prevent a Hamming weight leakage that would occur if clearing to zero. For secrets stored in multiple shares, use different permutations (or separate LFSRs) to perform the clearing of the shares.

Recommendation 3: Fault injection countermeasures

Consider defenses to fault injection / glitching attacks (survey and overview of attacks, see [12] and [13])

1. Initially assume that the adversary can glitch any node arbitrarily, and determine the resulting worst-case scenario. This is a very conservative approach and might lead to over-pessimism, but serves to highlight potential issues. Then, to ease implementation burden, assume the adversary can glitch to all-1's or all-0's (since these are considered "easier" to reach), and that reset can be asserted semi-arbitrarily.
2. Use parity/ECC on memories and data paths (note here that ECC is not true integrity, have to use hash to prevent forgery, see [1]). For memories, ECC is helpful to protect instruction streams or values that can cause "branching control flows" that redirect execution flow. Parity is potentially helpful if detection of corruption is adequate (though double-glitch fault injection can fool parity, so Hsiao or other detect-2-error codes can be used, even without correction circuitry implemented). When committing to an irrevocable action (e.g. burning into OTP, unlocking part of the device/increasing permissions), ECC is probably more appropriate.
   1. When selecting a specific ECC implementation, the error detection properties are likely more important than error correction (assuming memory lifetime retention/wear are not considered). For a good example of how to consider the effectiveness of error correction, see this PR comment.
3. State machines:
   1. Have a minimum Hamming distance for state machine transitions, to make single bit faults non-effective
   2. Use a sparsely populated state encoding, with all others marked invalid - see 11.1 about optimization concerns when doing this though
   3. All states could have the same Hamming weight, then can constantly check for this property (or use ECC-type coding on state variable and check this)
   4. If waiting for a counter to expire to transition to the next state, better if the terminal count that causes the transition is not all-0/all-1. One could use an LFSR instead of a binary counter, but debugging this can be a bit painful then
4. Maintain value-and-its-complement throughout datapath (sometimes called "dual rail" logic), especially if unlocking/enabling something sensitive, and continually check for validity/consistency of representation
5. Incorporate temporal randomness where possible (example: add delay cycles based on LFSR around sensitive operations)
6. Run-it-twice and compare results for sensitive calculations
7. Redundancy - keep/store multiple copies of sensitive checks/data
8. For maximum sensitivity, compare combinational and sequential paths with hair-trigger/one-shot latch of miscompare
9. Empty detection for OTP/flash (if needed, but especially for lifecycle determination)
10. Avoid local resets / prefer larger reset domains, since a glitch on this larger reset keeps more of the design "in sync." But, consider the implications of any block with more than one reset domain (see also 9.1).
11. Similar to the "mix in" idea of 4.4, in any case where multiple contributing "votes" are going to an enable/unlock decision, consider mixing them into some cryptographic structure over time that will be diverted from its path by attempts to glitch each vote. (Note: if the final outcome of this is simply a wide-compare that produces a single-bit final unlock/enable vote then this is only marginally helpful - since that vote is now the glitch target. Finding a way to bind the final cryptographic result to the vote is preferred, but potentially very difficult / impossible, depending on the situation.)
12. When checking/creating a signal to permit some sensitive operation, prefer that the checking logic is maximally volatile (e.g. performs a lot of the calculation in a single cycle after a register), such that a glitch prevents the operation. Whereas, when checking to deny a sensitive operation, prefer that the checking logic is minimally volatile (is directly following a register with minimal combinational logic), such that a glitch will be recovered on the next clock and the denial will be continued/preserved.
13. CFI (control flow integrity) hardware can help protect a processor / programmable peripheral from some types of glitch attacks. This topic is very involved and beyond the scope of these guidelines, consult [2] for an introduction to previous techniques.
14. Analog sensors (under/over-voltage, laser light, mesh breach, among others) can be used to generate SoC-level alerts and/or inhibit sensitive operations. Many of these sensors require calibration/trimming, or require hysteresis circuits to prevent false-positives, so they may not be usable in fast-reacting situations.
15. Running an operation (e.g. AES or KMAC) to completion, even with a detected fault, is sometimes useful since it suppresses information for the adversary about the success/failure of the attempted fault, and minimizes any timing side channel. However, for some operations (e.g. ECDSA sign), operations on faulty inputs can have catastrophic consequences. These guidelines cannot recommend a default-safe posture, but each decision about handling detected faults should be carefully considered.
16. For request-acknowledge interfaces, monitor the acknowledge line for spurious pulses at all times (not only when pending request) and use this as a glitch/fault detector to escalate locally and/or generate alerts.
17. When arbitrating between two or more transaction sources with different privilege/access levels, consider how to protect a request from one source being glitched/forged to masquerade as being sourced from another higher-privilege source (for example, to return side-loaded hardware-visible-only data via a software read path). At a minimum, redundant arbitration and multiple-bit encoding of the arbitration "winner" can help to mitigate this type of attack.

Recommendation 4: Handling of secrets

1. Diversify types/sources of secrets (e.g. use combination of RTL constants + OTP + flash) to prevent a single compromise from being effective
2. Rather than "check an unlock value directly" - use a hash function with a user-supplied input, and check the output of the hash matches. This way the unlock value is not contained in the netlist.
3. Qualify operations with allowed lifecycle state (even if redundant with other checks)
4. Where possible, mix in operating modes to calculation of derived secrets to create parallel/non-substitutable operating/keying domains. (i.e. mixing in devmode, lifecycle state)
   1. If defenses can be bypassed for debugging/recovery, considering mixing in activation vector/bypass bits of defenses as well, consider this like small-scale attestation of device state
5. Encrypt (or at least scramble) any secrets stored at-rest in flash/OTP, to reduce risks of static/offline inspection.

Recommendation 5: Alerts

1. Generate alerts on any detected anomaly (need to define what priority/severity should be assigned)
2. Where possible, prefer to take a local action (clearing/randomizing state, cease processing) in addition to generating the alert

Recommendation 6: Safe default values

1. All case statements, if statements, and ternaries should consider what the safest default value is. Having an "invalid" state/value is nice to have for this purpose, but isn't always possible.
2. Operate in a general policy/philosophy of starting with lowest allowed privilege and augmenting by approvals/unlocks.
3. Implement enforcement of inputs on CSRs - qualify/force data attempted to be written based on lifecycle state, peripheral state, or other values. The designer must determine the safest remapping, e.g. write --> read, read --> nop, write --> nop and so forth. Blanket implementation of input enforcement complicates verification, so this style of design should be chosen only where the inputs are particularly sensitive (requests to unlock, privilege increase requests, debug mode enables, etc).

Recommendation 7: DFT issues

1. Entry and exit from scan mode should cause a reset to prevent insertion or exfiltration of sensitive values
2. Ensure that when in production (e.g. not in lab debug) environments, scan chains are disabled
3. Processor debug paths (via JTAG) may need to be disabled in production modes
4. Beware of self-repair or redundant-row/columns schemes for memories (SRAM and OTP), as they can be exploited to misdirect reads to adversary-controlled locations

Recommendation 8: Power management issues

1. If module is not in an always-on power domain, consider that a sleep/wake sequence can be used to force a re-derivation of secrets needed in the module, as many times as desired by the adversary
2. Fine-grained clock gating should never be used for any module that processes secret data, only coarse-grained (module-level) gating is acceptable. (Fine grained gates essentially compute clock_gate = D ^ Q which often acts as an SCA "amplifier").

Recommendation 9: "Synchronization" (consistency) issues

1. If a module interacts with other modules in a stateful way (think of two data-transfer counters moving in ~lockstep, but the counts are not sent back and forth for performance optimization), what happens if:
   1. One side is reset and the other is not
   2. One side is clock-gated and the other is not
   3. One side is power-gated and the other is not
   4. The counter on one side is glitched
2. Generally these kind of blind lockstep situations should be avoided where possible, and current module/interface status should exchanged in both directions and constantly checked for validity/consistency

Recommendation 10: Recovery mechanism considerations

1. What happens if a security mechanism fails? (Classic problem of this variety is on-die sensors being too sensitive and resetting the chip) Traditionally, fuses can disable some mechanisms if they are faulty.
2. Could an adversary exploit a recovery mechanism? (If a sensor can be fuse-disabled, wouldn't the adversary just do that? See 4.4 above.)

Recommendation 11: Optimization concerns

1. Sometimes synthesis will optimize away redundant (but necessary for security) logic - dont_touch or size_only attributes may sometimes be needed, or even more aggressive preservation strategies. Example: when using the sparse FSM encoding, use the prim_flop component for the state vector register.
2. Value-and-complement strategies can also be optimized away, or partially disconnected such that only half of the datapath is contributing to the logic, or a single register with both Q & Qbar outputs becomes the source of both values to save area.
3. Retiming around pipeline registers can create DPA issues, due to inadvertent combination of shares, or intra-cycle glitchy evaluation. For DPA-resistant logic, explicitly declare functions and registers using prim_* components, and make sure that pipeline retiming is not enabled in synthesis.

Recommendation 12: Entropy concerns

1. Verify that all nonces are truly only used once
2. If entropy is broadcast, verify list of consumers and arbitration scheme to prevent reuse / duplicate use of entropy in sensitive calculations
3. Seeds for local LFSRs need to be unique/diversified

Recommendation 13: Global secrets

1. Avoid if at all possible
2. If not possible, have a process to generate/re-generate them; make sure this process is used/tested many times before final netlist; process must be repeatable/deterministic given some set of inputs
3. If architecturally feasible, install a device-specific secret to override the global secret once boot-strapped (and disable the global secret)

Recommendation 14: Sensors

1. Sensors need to be adjusted/tweaked so that they actually fire. It is challenging to set the sensors at levels that detect "interesting" glitches/environmental effects, but don't fire constantly or cause yield issues. Security team should work with the silicon supplier to determine the best course of action here.
2. Sensor configuration / calibration data should be integrity-protected.

Recommendation 15: Reset vs. non-reset flops

The flip-flop type used for registering sensitive assets inside a particular module should be selected later in the design process on a sub-block level and in consideration of all countermeasures and security hardening employed. For details, refer to Reset vs. non-reset flops

References and further reading

[1]: Overview of checksums and hashes - https://cybergibbons.com/reverse-engineering-2/checksums-hashes-and-security/

[2]: A Survey of hardware-based Control Flow Integrity - https://arxiv.org/pdf/1706.07257.pdf

[3]: Cache-timing attacks on AES - https://cr.yp.to/antiforgery/cachetiming-20050414.pdf

[4]: Meltdown: Reading Kernel Memory from User Space - https://meltdownattack.com/meltdown.pdf

[5]: Spectre Attacks: Exploiting Speculative Execution - https://spectreattack.com/spectre.pdf

[6]: Timing Attacks on Implementations of Diffie-Hellman, RSA, DSS, and Other Systems - https://www.rambus.com/wp-content/uploads/2015/08/TimingAttacks.pdf

[7]: Differential Power Analysis - https://paulkocher.com/doc/DifferentialPowerAnalysis.pdf

[8]: SoC it to EM: electromagnetic side-channel attacks on a complex system-on-chip - https://www.iacr.org/archive/ches2015/92930599/92930599.pdf

[9]: Introduction To differential power analysis - https://link.springer.com/content/pdf/10.1007/s13389-011-0006-y.pdf

[10]: Principles of Secure Processor Architecture Design - https://caslab.csl.yale.edu/tutorials/hpca2019/ and https://caslab.csl.yale.edu/tutorials/hpca2019/tutorial_principles_sec_arch_20190217.pdf

[11]: Time Protection - https://ts.data61.csiro.au/projects/TS/timeprotection/

[12]: Fault Attacks on Secure Embedded Software: Threats, Design and Evaluation - https://arxiv.org/pdf/2003.10513.pdf

[13]: The Sorcerer's Apprentice Guide to Fault Attacks - https://eprint.iacr.org/2004/100.pdf

[14]: Fault Mitigation Patterns - https://www.riscure.com/uploads/2020/05/Riscure_Whitepaper_Fault_Mitigation_Patterns_final.pdf

[15]: SIFA: Exploiting Ineffective Fault Inductions on Symmetric Cryptography - https://eprint.iacr.org/2018/071.pdf

Notes

Footnotes

1. In other OpenTitan documents, the combination of technical and administrative defense are often referred to as "logical security" ↩