FuzzBench: 2023-09-06-libafl-3 report

experiment summary

We show two different aggregate (cross-benchmark) rankings of fuzzers. The first is based on the average of per-benchmarks scores, where the score represents the percentage of the highest reached median code-coverage on a given benchmark (higher value is better). The second ranking shows the average rank of fuzzers, after we rank them on each benchmark according to their median reached code-covereges (lower value is better).

By avg. score

	average normalized score
fuzzer
libafl_fuzzbench_naive_ctx	97.48
honggfuzz	50.00
libafl_fuzzbench_ngram8	50.00
aflplusplus	49.75
afl	47.50
libfuzzer	46.51

By avg. rank

	average rank
fuzzer
honggfuzz	2.0
aflplusplus	2.5
libafl_fuzzbench_naive_ctx	2.5
afl	3.5
libafl_fuzzbench_ngram8	3.5
libfuzzer	4.0

Critical difference diagram

The diagram visualizes the average rank of fuzzers (second ranking above) while showing the significance of the differences as well. What is considered a "critical difference" (CD) is based on the Friedman/Nemenyi post-hoc test. See more in the documentation.

Note: If a fuzzer does not support all benchmarks, its ranking as shown in this diagram can be lower than it should be. So please check the list of supported benchmarks for the fuzzer(s) of your interest. The list could be specified in the fuzzer's README.md like this.

Median relative code-coverages on each benchmark

Note: The relative coverage summary table shows the median relative performance of each fuzzer to the experiment maximum. Thus the highest relative performance may not be 100%.
trial_relative_coverage = trial_coverage / experiment_max_coverage

	libafl_fuzzbench_ngram8	honggfuzz	libafl_fuzzbench_naive_ctx	aflplusplus	afl	libfuzzer
FuzzerMedian	99.00	94.00	94.00	93.00	89.00	87.00
FuzzerMean	99.00	94.00	94.00	93.00	89.00	87.00
brotli_decode_fuzzer	99.00	nan	99.00	nan	nan	nan
libaom_av1_dec_fuzzer	nan	94.00	89.00	93.00	89.00	87.00

Fuzzers are sorted by "FuzzerMean" (average median relative coverage), highest on the left.
Green background = highest relative median coverage.
Blue gradient background = greater than 95% relative median coverage.

brotli_decode_fuzzer summary

Ranking by median reached code coverage

Mean code coverage growth over time

* The error bands show the 95% confidence interval around the mean code coverage.

Sample statistics and statistical significance (code coverage)

Code coverage sample statistics

		count	mean	std	min	25%	median	75%	max
fuzzer	time
libafl_fuzzbench_ngram8	82800	15.0	899.533333	4.437932	891.0	896.5	901.0	903.0	904.0
libafl_fuzzbench_naive_ctx	82800	16.0	899.187500	3.673668	892.0	897.0	900.0	902.0	904.0

Vargha-Delaney A12 measure

The table summarizes the A12 values from the pairwise Vargha-Delaney A measure of effect size. Green cells indicate the probability the fuzzer in the row will outperform the fuzzer in the column.

Mann-Whitney U test

The table summarizes the p values of pairwise Mann-Whitney U tests. Green cells indicate that the reached coverage distribution of a given fuzzer pair is significantly different.

Unique code coverage plots

Ranking by unique code branches covered
Each bar shows the total number of code branches found by a given fuzzer. The colored area shows the number of unique code branches (i.e., branches that were not covered by any other fuzzers).

Pairwise unique code coverage
Each cell represents the number of code branches covered by the fuzzer of the column but not by the fuzzer of the row

Code coverage reports for each fuzzer on this benchmark

afl aflplusplus libafl_fuzzbench_naive_ctx libafl_fuzzbench_ngram8 libfuzzer