CCS Experiments
Use this before submission when the attack demonstration, defense evaluation, or measurement
story is not yet locked.
Experiment audit
- Map each security claim to a specific artifact: an exploit run, an overhead measurement, a
coverage number, a false-positive/false-negative table, or a measurement dataset.
- For attacks, demonstrate the exploit against a realistic, named target (software version,
platform, configuration) and report the resource cost to the attacker.
- For defenses, evaluate against an adaptive attacker built with knowledge of the defense,
and report performance overhead, memory cost, and any compatibility breakage.
- For measurements, validate sampling: document the population, the vantage point, coverage
and blind spots, and ground-truth checks against known cases.
- Include baselines that represent the state of the art in attack or defense, not strawmen.
- Report variance for stochastic results and audit for leakage, selection bias, and any
mismatch between the threat model and the tested configuration.
What experiments are for at this venue
- CCS experiments exist to make a security claim undeniable to a skeptic, not to top a
benchmark. One clean end-to-end exploit against a real target outweighs a table of
micro-benchmarks.
- The strongest defense design triad: the attack it stops, an adaptive attack that knows the
defense, and a deployment-cost measurement. Missing the middle element is the classic CCS
defense reject.
- Reviewers, often practitioners, check whether the evaluation environment matches the threat
model. A defense claimed for production but tested only on a toy in a lab invites the
relevance question.
Attack-and-defense evaluation table
| Security claim |
Matching evidence |
Reject pattern avoided |
| Exploit is practical |
End-to-end run on named target with attacker cost |
"Works only in a lab against a strawman" |
| Defense stops the attack |
Detection/prevention rate on the original attack |
"No numbers, only a design argument" |
| Defense resists adaptation |
Adaptive attacker with defense knowledge, degraded results |
"Only the non-adaptive attack was tried" |
| Deployment is feasible |
Overhead, memory, compatibility on a realistic workload |
"Security claimed, cost never measured" |
Vignette: evaluating a control-flow-integrity defense
Suppose the paper proposes a fine-grained CFI scheme. The matching plan: reproduce a known
code-reuse attack and show it blocked; construct an adaptive attacker that respects the CFI
policy and search for surviving gadget chains; then measure runtime overhead and binary-size
growth on a standard benchmark suite. Every claim ties to a numbered table, and the adaptive
result is reported even when it dents the headline.
Reporting floor
- Name every target's exact version and configuration; "a popular browser" is not a target.
- Report the attacker's resource budget (queries, time, samples) and the defense's measured
overhead rather than vague "negligible cost" language.
Output format
[Experiment readiness] strong / adequate / weak
[Claim -> evidence map] <claim: exploit run / overhead table / measurement>
[Missing security evidence] <adaptive attack / baseline / cost / validation>
[Threat-model mismatch] <where the setup breaks the stated model>
[Decision-critical next run] <one experiment>
1---2name: ccs-experiments3description: Use when designing or auditing ACM CCS experiments, attack demonstrations, adaptive-attack defense evaluations, security measurements, baselines, overhead and cost reporting, ablations, and claim-to-evidence fit, with emphasis on evidence that survives an adversarial program committee rather than leaderboard wins.4metadata:5 source: https://skillmd.com/skills/brycewang-stanford/ccs-experiments6---78# CCS Experiments910Use this before submission when the attack demonstration, defense evaluation, or measurement11story is not yet locked.1213## Experiment audit1415- Map each security claim to a specific artifact: an exploit run, an overhead measurement, a16 coverage number, a false-positive/false-negative table, or a measurement dataset.17- For attacks, demonstrate the exploit against a realistic, named target (software version,18 platform, configuration) and report the resource cost to the attacker.19- For defenses, evaluate against an adaptive attacker built with knowledge of the defense,20 and report performance overhead, memory cost, and any compatibility breakage.21- For measurements, validate sampling: document the population, the vantage point, coverage22 and blind spots, and ground-truth checks against known cases.23- Include baselines that represent the state of the art in attack or defense, not strawmen.24- Report variance for stochastic results and audit for leakage, selection bias, and any25 mismatch between the threat model and the tested configuration.2627## What experiments are for at this venue2829- CCS experiments exist to make a security claim undeniable to a skeptic, not to top a30 benchmark. One clean end-to-end exploit against a real target outweighs a table of31 micro-benchmarks.32- The strongest defense design triad: the attack it stops, an adaptive attack that knows the33 defense, and a deployment-cost measurement. Missing the middle element is the classic CCS34 defense reject.35- Reviewers, often practitioners, check whether the evaluation environment matches the threat36 model. A defense claimed for production but tested only on a toy in a lab invites the37 relevance question.3839## Attack-and-defense evaluation table4041| Security claim | Matching evidence | Reject pattern avoided |42|---|---|---|43| Exploit is practical | End-to-end run on named target with attacker cost | "Works only in a lab against a strawman" |44| Defense stops the attack | Detection/prevention rate on the original attack | "No numbers, only a design argument" |45| Defense resists adaptation | Adaptive attacker with defense knowledge, degraded results | "Only the non-adaptive attack was tried" |46| Deployment is feasible | Overhead, memory, compatibility on a realistic workload | "Security claimed, cost never measured" |4748## Vignette: evaluating a control-flow-integrity defense4950Suppose the paper proposes a fine-grained CFI scheme. The matching plan: reproduce a known51code-reuse attack and show it blocked; construct an adaptive attacker that respects the CFI52policy and search for surviving gadget chains; then measure runtime overhead and binary-size53growth on a standard benchmark suite. Every claim ties to a numbered table, and the adaptive54result is reported even when it dents the headline.5556## Reporting floor5758- Name every target's exact version and configuration; "a popular browser" is not a target.59- Report the attacker's resource budget (queries, time, samples) and the defense's measured60 overhead rather than vague "negligible cost" language.6162## Output format6364```text65[Experiment readiness] strong / adequate / weak66[Claim -> evidence map] <claim: exploit run / overhead table / measurement>67[Missing security evidence] <adaptive attack / baseline / cost / validation>68[Threat-model mismatch] <where the setup breaks the stated model>69[Decision-critical next run] <one experiment>70```