Memory Command Sequence Detection for Row Hammer Attacks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing memory systems, particularly DRAM, are vulnerable to row hammer attacks, which can cause data integrity issues due to excessive activation commands leading to inter-cell coupling and charge leakage, and current error correction techniques may not detect or correct multiple bit failures effectively.
Innovation Solution
Implementing an analog compute module with a dynamic random access memory and a non-volatile memory cell array programmed in a synapse mode to perform multiplication and accumulation operations, using an artificial neural network to classify anomalous command sequences and detect potential attacks by analyzing patterns of command usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If error correction code (ECC) techniques are used to detect and correct bit failures, then data integrity is improved, but multiple bit failures induced by row hammer attacks may not be detected or corrected effectively
Solution Approach 1:
The patent applies preliminary action by monitoring and analyzing command sequences before they can execute row hammer attacks. The system detects anomalous patterns of activation commands and prevents the actual memory operations that would cause charge leakage and bit failures, rather than attempting to correct errors after they occur.
Solution Approach 2:
The patent introduces an intermediary command sequence analyzer between the host and memory device. This intermediary component intercepts and analyzes commands before they reach the memory, identifying malicious patterns and blocking them, thus preventing the propagation of harmful command sequences that ECC cannot detect.
2Productivity
If activation commands are repeatedly applied to read DRAM, then data retrieval is enabled, but excessive activation commands cause inter-cell coupling and charge leakage leading to row hammer attacks
Solution Approach 1:
The patent implements feedback by continuously monitoring the sequence and frequency of activation commands. The system analyzes command patterns in real-time and provides feedback control by blocking anomalous command sequences that exhibit row hammer characteristics, thereby preventing excessive activation that would cause charge leakage while allowing normal operational commands to proceed.
Solution Approach 2:
The patent applies dynamics by making the command filtering process adaptive and dynamic. The command sequence analyzer learns and adapts to normal access patterns, dynamically adjusting its detection thresholds and blocking decisions based on the current operational context, enabling it to distinguish between legitimate high-frequency access and malicious row hammer attacks.
3Reliability
If an analog compute module with artificial neural network is implemented to detect anomalous command sequences, then security against row hammer attacks is improved, but device complexity increases
Solution Approach 1:
The patent replaces traditional mechanical or digital command filtering mechanisms with an analog compute module that uses continuous voltage signals to represent and process command sequences. The analog neural network processes command patterns through continuous voltage transformations rather than discrete digital logic, reducing latency and enabling real-time detection of malicious patterns without the complexity of digital signal processing hardware.
Data Source
AI summary
A device to detect attacks on a memory system in an advanced driver-assistance system (ADAS) of a vehicle. The device has an interface operable on a memory channel, a random access memory, a non-volatile memory cell array, and a controller configured to detect a trigger event, and in response: identify a sequence of commands received in the interface from the memory channel to access memory services provided via at least the random access memory during ADAS operations; perform operations of multiplication and accumulation using the non-volatile memory cell array to implement computations of an artificial neural network responsive to the sequence of commands as an input to generate a classification of the sequence as an output; and provide the classification via the interface.