Safety consideration: Neural Network 02
[Author]
Renhong WENG,
1.Artificial Intelligence safety investigator
2.Automotive hobbiest
First: why i pay attention to neural network safety analysis
1. Automated vehicle life killing accidents
Refer to long article 'CAST safety Analyze for Automated Vehicle Life-killling Accident' in https://mp.weixin.qq.com/s?__biz=MzUyNDc0Njc3Mw%3D%3D&mid=2247484578&idx=1&sn=b93046f8aab622ed7ffbf6915ed36644&scene=45#wechat_redirect, we absolutely know that there are some flaws in the mental/process model contributing to the actions, some of them relating to
| Component in Control Loop | Flaws contributing to the actions |
| Camera | Perception did not have good perception in the pedestrian or object, leading to disastrous reaction for automobile: - just 5.6 seconds before the collision, pedestrian are classified as object -just 4.6 seconds before the collision, pedestrian are classified as vehicle -just 2.6~2.5 seconds before the collision, pedestrian are classified as bicycle -just 1.6 seconds before the collision, pedestrian are classified as bicycle |
| LIDAR | |
| mmRADAR | |
| Sensor fusion model |
You can see the camera had as well some problems in the neural network, and even in LIDAR, mmRADAR, still some problems in the deep neural network which will be life-threatening.
2. EU GDPR requirements
In Recital 71 Profiling, there are some direct personal datas if they had been put into automated processing, treated, they have to be explainable and trusted by the data owner, detail requirements as following:
In any case, such processing should be subject to suitable safeguards, which should include specific information to the data subject andthe right to obtain human intervention, to express his or her point of view, to obtain an explanation ofthe decision reached after such assessment and to challenge the decision. Suchmeasure should not concern a child.
However,neural network sometimes can beunexpainable, so that if personal data used and been NNed, which violates theGDPR.
Second: simple example
Third: safety properties of Neural Network
Forth: example codings
# Import Data
df = pd.read_csv("https://raw.githubusercontent.com/selva86/datasets/master/mpg_ggplot2.csv")
df_select = df.loc[df.cyl.isin([4,8]), :]
# Plot
sns.set_style("white")
gridobj = sns.lmplot(x="displ", y="hwy", hue="cyl", data=df_select, height=7, aspect=1.6, robust=True, palette='tab10', scatter_kws=dict(s=60, linewidths=.7, edgecolors='black'))
# Decorations
gridobj.set(xlim=(0.5, 7.5), ylim=(0, 50))plt.title("Scatterplot with line of best fit grouped by number of cylinders", fontsize=20)
Source: https://mp.weixin.qq.com/s/cXVg1Ez-qbs-gIs2FDyaRg
Thanks for your reading, and if you feel good, please have some comments for me for progress.
[REF]
1.https://mp.weixin.qq.com/s?__biz=MzUyNDc0Njc3Mw%3D%3D&mid=2247484578&idx=1&sn=b93046f8aab622ed7ffbf6915ed36644&scene=45#wechat_redirect
2.Safe and Trustworthiness of Deep Neural Networks: A survey
3.GDPR
4.https://mp.weixin.qq.com/s/cXVg1Ez-qbs-gIs2FDyaRg
- 用户评论