模型鲁棒性增强 - GAN 对抗样本生成

GAN 配置

数据采集方式:通过调整噪声水平和迭代次数,控制GAN生成对抗样本的强度。数据通过输入框手动调整,点击“生成对抗样本”按钮模拟生成过程,实际应用中会调用后端服务。

模型训练配置

数据采集方式:通过手动输入模型名称、学习率和批大小,模拟模型训练过程。实际应用中会将这些配置传递给后端进行模型训练。

模型鲁棒性评估

数据采集方式:点击“评估模型”按钮,模拟模型鲁棒性评估过程。实际应用中会调用后端服务进行多种对抗攻击测试,并将结果返回到前端展示。

测试数据:

攻击类型: FGSM

成功率: 10%

攻击类型: PGD

成功率: 25%

攻击类型: DeepFool

成功率: 15%

差分隐私设置

数据采集方式:通过手动调整Epsilon和Delta参数,模拟差分隐私的应用。实际应用中会将这些参数传递给后端,用于在训练过程中添加噪声,保护数据隐私。

SQL 查询

数据生成方式:SQL 查询用于从数据库中检索特定信息,例如GAN数据的质量指标或Corner Case的覆盖度。 这些查询可以帮助评估GAN的性能和数据质量。

SQL 数据库查询语句

数据生成方式:SQL 查询用于从数据库中检索特定信息,例如GAN数据的质量指标或Corner Case的覆盖度。 这些查询可以帮助评估GAN的性能和数据质量。

-- 修改 model_evaluations 表

ALTER TABLE `autonomous_driving`.`model_evaluations` ADD COLUMN `sensor_data_id` BIGINT UNSIGNED NULL COMMENT '外键,关联`sensor_data`表' AFTER `model_id`, ADD COLUMN `gan_id` BIGINT UNSIGNED NULL COMMENT '外键,关联`gan_metadata`表' AFTER `sensor_data_id`, ADD INDEX `fk_model_evaluations_sensor_data_idx` (`sensor_data_id` ASC) VISIBLE, ADD INDEX `fk_model_evaluations_gan_metadata_idx` (`gan_id` ASC) VISIBLE, ADD CONSTRAINT `fk_model_evaluations_sensor_data` FOREIGN KEY (`sensor_data_id`) REFERENCES `autonomous_driving`.`sensor_data` (`sensor_data_id`) ON DELETE CASCADE ON UPDATE CASCADE, ADD CONSTRAINT `fk_model_evaluations_gan_metadata` FOREIGN KEY (`gan_id`) REFERENCES `autonomous_driving`.`gan_metadata` (`gan_id`) ON DELETE CASCADE ON UPDATE CASCADE;

-- 视频流预处理实时性视图

CREATE OR REPLACE VIEW `autonomous_driving`.`view_preprocessing_throughput` AS SELECT sd.sensor_data_id, sd.timestamp AS frame_timestamp, dp.process_timestamp AS preprocessing_timestamp, TIMESTAMPDIFF(MICROSECOND, sd.timestamp, dp.process_timestamp) / 1000000 AS preprocessing_latency_seconds FROM `autonomous_driving`.`sensor_data` sd JOIN `autonomous_driving`.`data_provenance` dp ON sd.sensor_data_id = dp.sensor_data_id WHERE dp.process_name = '视频预处理';

-- 数据质量监控视图 (GAN 数据质量)

CREATE OR REPLACE VIEW `autonomous_driving`.`view_gan_data_quality` AS SELECT sd.sensor_data_id, gm.gan_id, me.metric_name, me.metric_value, me.evaluation_timestamp FROM `autonomous_driving`.`sensor_data` sd JOIN `autonomous_driving`.`gan_metadata` gm ON sd.sensor_data_id = gm.gan_id -- 修改了 ON 条件 JOIN `autonomous_driving`.`model_evaluations` me ON gm.gan_id = me.gan_id WHERE sd.source = 'GAN' AND me.metric_name IN ('KL散度', 'JS散度', 'Wasserstein距离', '置信度', '人工评分');

-- 融合模块集成测试视图

CREATE OR REPLACE VIEW `autonomous_driving`.`view_integration_test_results` AS SELECT sd.sensor_data_id, sd.scene, pmp.model_name, pmp.parameter_name, pmp.parameter_value, pmp.units FROM `autonomous_driving`.`sensor_data` sd JOIN `autonomous_driving`.`physical_model_parameters` pmp ON sd.sensor_data_id = pmp.sensor_data_id WHERE sd.source = 'synthetic' AND pmp.model_name IN ('最小安全距离', '最大允许速度', '碰撞时间');

-- Corner Case 场景覆盖度分析视图

CREATE OR REPLACE VIEW `autonomous_driving`.`view_corner_case_coverage` AS SELECT ccd.corner_case_id, ccd.name, ccd.risk_level, ccd.severity_score FROM `autonomous_driving`.`corner_case_definitions` ccd;

-- GAN 结构验证视图

CREATE OR REPLACE VIEW `autonomous_driving`.`view_gan_structure_validation` AS SELECT gm.gan_id, gm.gan_type, me.metric_name, me.metric_value, me.evaluation_timestamp FROM `autonomous_driving`.`gan_metadata` gm JOIN `autonomous_driving`.`model_evaluations` me ON gm.gan_id = me.gan_id WHERE me.metric_name IN ('Inception Score', 'Fréchet Inception Distance', 'Perceptual Path Length');

-- 数据偏见检测视图

CREATE OR REPLACE VIEW `autonomous_driving`.`view_data_bias` AS SELECT s.sensor_data_id, s.data_type, dbd.bias_type, dbd.bias_score, dbd.detection_method FROM autonomous_driving.sensor_data s JOIN autonomous_driving.data_bias_detection dbd ON s.sensor_data_id = dbd.sensor_data_id;

-- 数据溯源视图

CREATE OR REPLACE VIEW `autonomous_driving`.`view_data_lineage` AS SELECT sd.sensor_data_id, sd.data_type, dp.process_name, dp.process_timestamp, dp.parameters FROM `autonomous_driving`.`sensor_data` sd JOIN `autonomous_driving`.`data_provenance` dp ON sd.sensor_data_id = dp.sensor_data_id;