THE WIRE · UPDATED 8:14 PM MDT No story is too small.
NewsPlain City

Research Examines Bit-Flip Sensitivity in Smallsat Neural Networks

Official source This story comes from an official government or institutional source.
A research presentation addressed the risks of bit-flip sensitivity in neural networks used by smallsat missions. The work explored memory corruption issues and the standard responses to such faults in flight software.

Key takeaways

  • Smallsat missions are increasingly deploying neural networks on COTS processors.
  • A single event upset caused by a charged particle can flip a single bit of memory.
  • Memory upsets in network weights can lead to incorrect values without raising a fault condition.
  • Standard cFS responses to memory corruption involve process restarts that discard runtime state.

Research presented at the Salt Palace Convention Center in Salt Lake City examined how single event upsets affect neural networks on smallsat missions. As these missions increasingly use COTS processors, a charged particle striking a device can flip a single bit of memory.

According to the research, an upset in a network weight does not raise a fault condition. Instead, the network continues to operate and returns incorrect values, which can lead to downstream control loops acting on incorrect data. The standard response within cFS to memory corruption is a process restart, a method that interrupts service and discards runtime state.

The study, titled Bit-Flip Sensitivity and Selective FP32 Protection for Onboard Neural Networks in Flight Software, focused on in-place SECDED healing of an FP32 weight buffer inside NASA cFS.

Article details

CategoryNews
CityPlain City
ToneNeutral
SourceAI Generated