The leakage taxonomy#
Seven channels, seven disclosures, seven separate arguments. This week adds no eighth. Which of them would you close first, and what measurement would defend the choice?
The six axes#
- medium — which physical quantity carries it out
- proximity — how close, and for how long
- availability — during the computation, or only after
- rate — bits of the secret per measurement
- volume — measurements needed before the estimate settles
- cost to close — what the defender gives up
Four you can fill in from your own Bench notes. Two you cannot.
“It leaks” is not a number#
Every channel so far was shown to leak something. None of those demonstrations said how much, in bits, under realistic noise.
Rate and volume are that gap.
Mutual information, briefly#
For a secret S and a measurement M of it:
I(S; M) = H(S) - H(S | M)How much uncertainty about S remains, on average, once you know M. Zero
means the channel told you nothing. H(S) means it told you everything.
On too few samples it reports a leak that is not there. Recompute it as the sample count grows and watch where it stops moving.
Placing the channels#
| Channel | Proximity | Availability | Rate |
|---|---|---|---|
| Acoustic (wk 2) | same room | during | low, many needed |
| Power (wk 3) | physical contact | during | high, few traces |
| Timing (wk 4) | network path | during | very low, very many |
| Cache (wk 6) | same package, any core, shared pages | during | medium |
| Thermal (wk 8) | none during, camera after | after only | low, decays |
Electromagnetic (wk 5) and optical (wk 7) are yours to place. There is no single correct table — the Bench and the capstone both test the reasoning, not the entries.
Two orderings, not one#
Rank the seven by cost to the attacker. Now rank them again by cost to the defender.
The lists disagree.
- cheap to exploit, cheap to close → a bug, already fixed
- costly to exploit, costly to close → someone else’s threat model
- cheap to exploit, costly to close → the ones that matter
A number a defender can act on#
“0.2 bits per trace, needs physical contact” and “6 bits per trace, over the network” are not the same engineering problem, even though both are, technically, side channels.
Bring your seven-row table and both orderings to the Bench.